Clay Millennium Prize problem (stated 2000, $1,000,000) · fluid dynamics / nonlinear PDE · status: OPEN
How to read this page. The unknown is a velocity field u(x,t) = (u1,u2,u3) and a scalar pressure p(x,t), with x ∈ ℝ3 (or the torus 𝕋3) and time t ≥ 0. We write ∂t for the time derivative, ∇ = (∂1,∂2,∂3) for the spatial gradient, Δ = ∂12+∂22+∂32 for the Laplacian, and ∇·u = ∂1u1+∂2u2+∂3u3 for the divergence. The constant ν > 0 is the kinematic viscosity. The symbol ω = ∇×u is the vorticity (the local spin of the fluid). "Smooth" means C∞ in space and time; "blow-up" means some derivative becomes infinite at a finite time. Throughout, regularity is the technical word for "stays smooth," and the whole problem is a fight to keep regularity from being destroyed by the equation's nonlinearity.
0 · The claim, in one line
Start the three-dimensional incompressible Navier–Stokes equations from any smooth, divergence-free, finite-energy velocity field and let viscosity act. Does the solution remain smooth for all time, or can it spontaneously develop a singularity (infinite velocity or infinite velocity-gradient) at some finite time? No one knows. Proving either alternative, rigorously, wins the Clay Millennium Prize. The problem has been open since (and long before) its formal posting in 2000.
1 · The object: what it is and why anyone cares
Navier–Stokes is the equation of a fluid. Written down by Claude-Louis Navier in the 1820s and put on firm footing by George Gabriel Stokes in the 1840s, it is Newton's second law (mass × acceleration = force) applied to an infinitesimal parcel of an incompressible viscous fluid (water, roughly). Every quantitative prediction about water flowing in a pipe, air over a wing, blood in an artery, ocean currents, or the weather is, at bottom, a computation with this equation. It is arguably the most-used partial differential equation in the physical sciences.
So the "why we care" is immediate on the engineering side. But the mathematical anxiety is deeper, and it is worth chasing to its root. We solve Navier–Stokes numerically millions of times a day and the answers match reality beautifully. Yet we cannot prove that the equation we are solving even has a solution (a single, well-defined, smooth answer) for all time in three dimensions. We do not know whether the model that underpins all of fluid engineering is internally consistent as a piece of mathematics. If a smooth flow of ordinary water, stirred by ordinary smooth forces, could concentrate its energy into an infinitely sharp spike in finite time, that would be a catastrophic failure of the model: and it would be the mathematical signature of something physicists have chased for a century: turbulence. Finite-time blow-up, if it exists, is where the smooth continuum description of a fluid tears itself apart. The Millennium Problem asks, with full rigor, whether that tear can happen.
The root of the "why," then, is this: Navier–Stokes is the simplest classical field theory whose fundamental question: does a solution exist?: is genuinely unanswered. It is a stress test of whether the mathematics of nonlinear partial differential equations is mature enough to describe the physical world it already models in practice.
2 · The precise statement / the mathematical setup
For an incompressible fluid of constant density (normalized to 1) filling all of ℝ3, with no external forcing, the equations are:
∂tu + (u·∇)u = −∇p + νΔu ∇·u = 0Top: Newton's law for a fluid parcel: the rate of change of momentum (left) equals the pressure force, plus the viscous friction force νΔu. Bottom: incompressibility: the fluid is neither created nor destroyed.
Here (u·∇)u is the convective (inertial) term, the only nonlinearity: component i is ∑j uj ∂j ui. The pressure p is not an independent unknown: it is the Lagrange multiplier that enforces ∇·u = 0, determined instantaneously from u by the (nonlocal) elliptic relation Δp = −∑i,j ∂iuj ∂jui. This nonlocality is one of the structural headaches.
The initial data u(x,0) = u0(x) is prescribed: smooth, divergence-free, and rapidly decaying (so the total energy ∫ℝ³|u0|2 dx is finite). The official Clay problem, formulated by Charles Fefferman, asks you to prove one of the following:
(A) / (B): Global existence and smoothness. On ℝ3 (A) or on the periodic torus 𝕋3 (B), for every such u0 there exist smooth u, p solving the equations for all t > 0 with bounded energy; or
(C) / (D): Breakdown. Exhibit smooth data on ℝ3 (C) or 𝕋3 (D) for which no such global smooth solution exists.
Proving any one of the four settles the prize. The analogous statements with a smooth forcing termf added to the momentum equation are also acceptable: a subtlety that matters later.
Two a-priori facts are the entire budget of "free" information we get. First, the energy inequality, obtained by multiplying the equation by u and integrating:
½‖u(t)‖L²2 + ν∫0t‖∇u(s)‖L²2 ds ≤ ½‖u0‖L²2Kinetic energy plus total viscous dissipation up to time t is no more than the initial energy. Energy can only decrease. This is essentially the ONLY coercive, globally-valid bound we possess.
Second, the scaling symmetry: if u(x,t) solves the equations, then so does
uλ(x,t) = λ u(λx, λ2t), pλ(x,t) = λ2 p(λx, λ2t)Zooming in by a factor λ in space, λ² in time, and λ in amplitude maps solutions to solutions. Remember this symmetry: the whole difficulty lives here.
3 · Every connection: what hangs on it and what it unlocks
Navier–Stokes is not an island. A resolution would ripple outward:
The theory of turbulence. Kolmogorov's 1941 statistical theory, the energy cascade, intermittency: all presuppose a solution theory they cannot rigorously stand on. Finite-time blow-up would be the exact mechanism by which energy reaches arbitrarily small scales. A regularity proof, conversely, would pin down the smallest scale viscosity permits.
The Euler equations (ν = 0, no viscosity). Navier–Stokes is the viscous regularization of Euler. Blow-up for 3-D Euler is also open and is widely believed to be the easier, more likely place for a singularity to live: viscosity helps, it does not hurt. Progress on one informs the other. (Elgindi proved finite-time singularity for 3-D Euler with low-regularity C1,α data in 2021: but that is Euler, not Navier–Stokes, and not smooth data.)
Critical-space well-posedness and harmonic analysis. The problem forced the development of a whole machinery (Besov spaces, BMO−1, paraproducts, maximal-regularity estimates) that now pervades the study of dispersive and parabolic PDE far beyond fluids.
The limits of "energy methods." As we will see, Navier–Stokes became the canonical example where the standard toolkit of PDE (energy estimates, Sobolev embeddings) provably cannot reach the goal. That realization reshaped how analysts think about supercritical equations generally.
Uniqueness and determinism. Does the equation predict a unique future from a given present? Recent work (below) has shown the honest answer, at the level of weak solutions, is unsettling: and that bears directly on whether Navier–Stokes is even a deterministic model.
4 · Every serious attack: and exactly how far each got
4.1 · Leray's weak solutions (1934): existence, but of what?
Jean Leray's 1934 paper (Acta Mathematica) is the foundation stone. Unable to produce smooth solutions, Leray lowered his sights and built weak (he called them "turbulent") solutions: objects satisfying the equations only in an averaged, integrated sense, living in the energy space. He proved such a weak solution exists globally in time for any finite-energy datum, and satisfies the energy inequality. Eberhard Hopf extended the construction to bounded domains in 1951; these are now the Leray–Hopf weak solutions. The catch: Leray could not prove his weak solutions are smooth, nor that they are unique. Ninety years later, in 3-D, we still cannot. Everything since is an attempt to close the gap between "a weak solution exists" and "the weak solution is smooth and unique."
4.2 · The two-dimensional case: completely solved
In two dimensions the problem is done: global existence, uniqueness, and smoothness all hold, for arbitrarily large data. This was established through the work of Jean Leray and Olga Ladyzhenskaya (whose 1960s monograph is definitive), with uniqueness via Lions–Prodi (1959). The reason 2-D works is sharp and instructive, and it is the single best guide to why 3-D is hard: see §5.
If you cannot prove smoothness outright, prove conditional smoothness: "if the solution satisfies bound X, then it is smooth." The classical result (Prodi 1959, Serrin 1962, Ladyzhenskaya 1967) is that a Leray–Hopf solution is smooth on (0,T] if it lies in LptLqx with
2/p + 3/q = 1, 3 < q ≤ ∞A "scaling-invariant" (critical) integrability threshold: cross it and the solution cannot be singular.
The endpoint q = 3 (i.e. u ∈ L∞tL3x) resisted until Escauriaza, Seregin and Šverák (2003) proved even it forces regularity. A closely related gauge is the Beale–Kato–Majda criterion (1984): a smooth solution survives past time T if and only if ∫0T‖ω(·,t)‖L∞ dt < ∞: blow-up requires the vorticity to become non-integrably large. These criteria are genuine theorems and immensely useful, but each merely relocates the mystery: no one can prove the hypotheses hold for general data.
4.4 · Small-data global existence: Fujita–Kato and critical spaces
If the data is small (in the right scale-invariant norm), the nonlinearity is dominated by the dissipation and global smooth solutions exist. Fujita–Kato (1964) proved this in the critical Sobolev space Ḣ1/2; Kato (1984) in L3; and Koch–Tataru (2001) in the largest known critical space BMO−1. These results are sharp and beautiful: but smallness is the whole game. The Millennium Problem is precisely about large data, where the nonlinearity wins locally and the question is whether it can win catastrophically.
4.5 · Partial regularity: Scheffer and Caffarelli–Kohn–Nirenberg (1976–82)
If you cannot show the singular set is empty, bound its size. Vladimir Scheffer (in a series of papers from 1976) initiated the measure-theoretic study of the possible singular set of a suitable weak solution. The crowning result is Caffarelli–Kohn–Nirenberg (1982):
𝓗1par(S) = 0For a "suitable" weak solution, the set S of space-time singular points has one-dimensional parabolic Hausdorff measure zero: in particular its parabolic Hausdorff dimension is at most 1.
In plain terms: singularities, if they exist at all, cannot form on a surface or even a curve of positive length in space-time: they are confined to an exceptionally thin set. Fanghua Lin (1998) gave a streamlined proof. This is the deepest unconditional structural result we have, and it is still a country mile from the goal: dimension "at most 1" is not "empty." (Note: it remains unknown whether suitable weak solutions can have even a single singular point.)
4.6 · The barrier results: energy methods proven insufficient (2016)
Terence Tao's 2016 paper (J. Amer. Math. Soc.) is a different kind of attack: instead of trying to solve the problem, prove that a whole class of methods cannot. Tao constructed an averaged Navier–Stokes equation (same scaling, same energy identity, same basic structure as the real equation, differing only in the fine details of how the nonlinearity mixes frequencies) and proved that this averaged equation does blow up in finite time. The mechanism: energy is handed from one scale to a smaller scale, then a smaller one, faster and faster, a self-amplifying cascade that reaches scale zero in finite time. See §5 for why this is devastating.
4.7 · Non-uniqueness via convex integration (2019–2022)
Tristan Buckmaster and Vlad Vicol (2019, Annals of Mathematics), adapting the convex-integration technology Camillo De Lellis and László Székelyhidi built for Euler, proved that weak solutions of 3-D Navier–Stokes are not unique: below the Leray–Hopf energy class, you can build infinitely many wild weak solutions with the same data. Albritton, Brué and Colombo (2022, Annals) pushed non-uniqueness up into the Leray–Hopf class itself, for the equation with a forcing term. This does not resolve the smooth problem: but it tells us the surrounding weak theory is far more pathological than one would hope, and that uniqueness and regularity are deeply intertwined.
5 · Why it breaks down: the honest crux
Wall 1: Supercriticality: the one scaling exponent that dooms the energy method. This is the central wall; everything else is downstream. Our only global bound is the energy, controlling ‖u‖L². Now feed the energy through the scaling symmetry uλ(x,t) = λu(λx,λ2t). A computation gives ‖uλ‖L²(ℝ³) = λ−1/2‖u‖L². As we zoom into small scales (λ → ∞), the energy norm of the rescaled solution goes to zero. The quantity we control becomes weaker and weaker exactly where a singularity would form: at infinitesimally small scales. In the language of PDE this makes Navier–Stokes supercritical with respect to the energy in 3-D. The scale-invariant norms that would actually control the nonlinearity (Ḣ1/2, L3, BMO−1) are all stronger than energy, and nothing keeps them bounded for large data. There is a definite gap between the half-derivative of control a bounded solution would need and the zero net derivatives the energy supplies: and that gap does not close. In 2-D the same computation gives ‖uλ‖L²(ℝ²) = ‖u‖L²: the energy is scale-invariant (critical), it controls the nonlinearity at every scale, and the problem is solved. One exponent: −1/2 versus 0: is the entire difference between a ninety-year-old open problem and a textbook theorem.
The supercriticality gap. In 3-D the only quantity we can globally control (the energy) sits strictly on the weak side of the critical line that would tame the nonlinearity. In 2-D the energy lands exactly on that line: and so do the 3-D critical norms (L³, Ḣ^½, BMO⁻¹), which by definition are scale-invariant; the trouble in 3-D is that nothing keeps them bounded for large data.
Wall 2: Vortex stretching: the term that exists only in 3-D. Take the curl of the momentum equation to get the vorticity equation: ∂tω + (u·∇)ω = (ω·∇)u + νΔω. Every term here also appears in 2-D: except(ω·∇)u, the vortex-stretching term. In two dimensions vorticity points out of the plane, is effectively a scalar, and (ω·∇)u = 0 identically; the vorticity is merely transported and diffused, so ‖ω‖ cannot grow, and by Beale–Kato–Majda the flow stays smooth. In three dimensions (ω·∇)u is a genuine feedback: the velocity field stretches and tilts vortex tubes, and a stretched tube spins faster (conservation of angular momentum), which intensifies the velocity gradients, which stretches the tube still more. This is the physical engine that could, in principle, drive ‖ω‖L∞ to infinity in finite time. No a-priori estimate is known to rule it out. The vortex-stretching term is not a technicality: it is the concrete, physical face of the abstract supercriticality in Wall 1.
Wall 3: Energy methods are provably not enough (the Tao barrier). One might hope some clever energy-type estimate: some positive quantity that the equation forces to stay bounded: would eventually crack the problem. Tao's 2016 averaged equation (§4.6) kills that hope for a large, natural class of approaches. He built an equation sharing Navier–Stokes' exact energy identity, scaling, and quadratic structure, differing only in the signs/weights of how Fourier modes interact: and proved it blows up in finite time by an explicit self-amplifying cascade across scales. The lesson is brutal: any proof of global regularity must use some finer feature of the true nonlinearity than its energy, scaling, and algebraic type, because a nearby equation with all three of those features still blows up. Every purely energy-based strategy is, by this result, dead on arrival. This does not say the true equation blows up: the true nonlinearity has special cancellation structure (divergence form, the precise geometry of vortex stretching) the averaged one lacks. But it means a winning proof has to exploit that structure in a way no one has yet found. It is a signpost reading "not this way," planted across the widest road.
Wall 4: The partial-regularity ceiling can't reach dimension zero. Caffarelli–Kohn–Nirenberg squeeze the singular set down to parabolic Hausdorff dimension ≤ 1 (§4.5). The method is an ε-regularity theorem: if a certain scale-invariant local quantity is small on a parabolic cylinder, the solution is smooth at its center. Covering arguments then bound the set where it fails to be small. But the scaling of the available local energy quantities bottoms out at dimension 1: the same supercritical half-derivative gap of Wall 1 reappears as a loss in the covering estimate, and there is no known local quantity whose scaling would push the bound below 1 down to the 0 (emptiness) that regularity requires. Forty years of effort have refined the constants and simplified the proofs but have not moved the dimension exponent. The ceiling is structural, not a lack of cleverness.
Wall 5 (FRINGE / speculative): engineered blow-up: self-similarity, and the "fluid computer."This box surveys speculative directions and labels them as such. The most natural blow-up ansatz is Leray's own: a self-similar singularity, u(x,t) ≈ (T−t)−1/2 U(x/√(T−t)), a profile that simply rescales as it collapses. This concrete scenario was ruled out: Nečas–Růžička–Šverák (1996) showed no nontrivial Leray self-similar solution exists in L3, and Tsai (1998) extended the exclusion. Attention moved to discretely self-similar and more exotic profiles, still open and actively studied (Chae, Wolf, and others).
A genuinely unconventional program, associated with Tao, reframes blow-up as computation: if one could embed a self-replicating logical "machine" into a fluid flow: a configuration that, using the equation's own dynamics, builds a half-scale copy of itself in a fraction of the time, forever: the result would be a finite-time cascade to zero scale, i.e. blow-up. The plausibility of this got a real boost when Cardona, Miranda, Peralta-Salas and Presas (2021) constructed Turing-complete steady Euler flows on certain 3-manifolds: fluid flows that can, in principle, simulate any computer. Why it hasn't closed the problem: those constructions are for Euler (no viscosity), often on curved or compact manifolds, and in steady rather than self-amplifying time-dependent form. Transplanting a universal constructor into viscous Navier–Stokes on flat ℝ3: where viscosity relentlessly smooths out small scales and fights exactly the miniaturization the machine needs: is an unsolved engineering problem of staggering difficulty, and may be impossible precisely because viscosity wins. It is a program, an idea with one real theorem behind it, not a result on the actual equation. Treat it as an honest frontier, not a near-miss.
6 · Status: confirmed open
The problem is open. As of this writing no one has proven global regularity, and no one has exhibited a finite-time singularity, for the 3-D incompressible Navier–Stokes equations with smooth finite-energy data. The claim is routinely checked and remains unresolved; the Clay Millennium Prize stands unclaimed.
What is honestly, rigorously settled:
Global weak (Leray–Hopf) solutions exist in 3-D: but not known to be smooth or unique.
The 2-D problem is completely solved (global, smooth, unique).
Local-in-time smooth solutions exist for any smooth data (short-time well-posedness), and global smooth solutions exist for small data in critical spaces.
The singular set, if any, has parabolic Hausdorff dimension ≤ 1 (CKN).
Many conditional regularity criteria (Prodi–Serrin, Beale–Kato–Majda, the L3 endpoint) are proven: none verifiable for general data.
Pure energy methods are provably insufficient (Tao); weak solutions are non-unique (Buckmaster–Vicol; Albritton–Brué–Colombo with forcing).
What a genuine proof would still require: for regularity, a global a-priori bound in some norm strictly stronger than energy: one that survives the supercritical scaling of Wall 1, controls the vortex-stretching feedback of Wall 2, and exploits a structural feature of the true nonlinearity that the Tao-averaged equation of Wall 3 lacks. For blow-up, an explicit (or implicit-but-rigorous) smooth initial datum together with a proof that its solution's velocity-gradient escapes to infinity in finite time, surviving the smoothing power of viscosity on flat space: defeating the obstructions that sank the self-similar scenario of Wall 5. Either direction demands a genuinely new idea, not a refinement of existing estimates.
Credit where it is due. The architecture of this subject was built by Jean Leray (1934, weak solutions) and Eberhard Hopf (1951); the two-dimensional theory by Leray, Olga Ladyzhenskaya, and Lions–Prodi; the conditional regularity theory by Giovanni Prodi (1959), James Serrin (1962), and Ladyzhenskaya (1967), with the endpoint by Escauriaza, Gregory Seregin and Vladimír Šverák (2003) and the vorticity criterion by Beale, Kato and Majda (1984); small-data theory by Fujita and Kato (1964), Kato (1984), and Koch–Tataru (2001); partial regularity by Vladimir Scheffer (from 1976) and Luis Caffarelli, Robert Kohn and Louis Nirenberg (1982), simplified by Fanghua Lin (1998); the insufficiency of energy methods by Terence Tao (2016); non-uniqueness by Tristan Buckmaster and Vlad Vicol (2019) and by Dallas Albritton, Elia Brué and Maria Colombo (2022); the exclusion of self-similar blow-up by Nečas, Růžička and Šverák (1996) and Tsai (1998); and the official problem statement by Charles Fefferman for the Clay Mathematics Institute (2000). The unsolved problem is claimed by no one here: if it is ever solved, that proof is offered freely to the world.