Page 1 of 1

"The" Navier-Stokes Problem

Posted: Sat Sep 12, 2026 3:26 pm
by Physics Guy
A recent news headline is about an AI solution to "the Navier-Stokes Problem", which is one of the seven Millennium Prize Problems defined in 2000 by the Clay Mathematics Institute. This is an interesting development, but not nearly as good as it might sound.

The Navier-Stokes equations are the basic equations of fluid dynamics. They describe how the velocity field in a fluid—wind speed and direction, in air, or current speed and direction in water, for example—must change over time, given how it is at any one time. The NS equations don't only apply to water and air; in principle they can govern any kind of flowing substance, whether liquid or gaseous or even in-between. These equations include the effects of viscosity and of pressure, but they don't define or explain either of them; the fluid's viscosity has to just be known in advance, through some other means, and the Navier-Stokes equations have to be supplemented with some other equations relating the fluid's pressure to things like its density and temperature. Figuring out those things, however, is usually much easier than the big step of solving the Navier-Stokes equations for given pressure and viscosity laws.

The Navier-Stokes equations are really tough to solve. Depending on how one looks at it, the fact that the Navier-Stokes equations are so hard to solve is either a reflection of, or a cause of, the fact that wind and water motion can be so unpredictable. The most obvious meaning of "solving the Navier-Stokes problem" is "finding an efficient way of solving the Navier-Stokes equations in any circumstances". That would be really great, but that is not what has just been done.

In this case, "the" Navier-Stokes problem just means the particular one of the seven Clay Institute problems that has to do with the Navier-Stokes equations. What it has to do with them is something very much less impressive than solving them completely. Instead it's a very specific technical question about whether a certain kind of velocity field can ever appear under a certain class of circumstances. This very specific question could be answered either by proving that, no, that kind of flow can never develop, or else by finding any one specific example in which the equations do predict that kind of thing happening.

This AI has found such an example, so the Clay Institute's specific problem is solved. In contrast to the logical difficulty of constructing an airtight general proof, finding an example like this is obviously the kind of thing than an AI might do better than humans could, just by brute-force searching. And in this case there seems to be some controversy over whether the AI mainly just copied some recent work by one group of human mathematicians and then filled in the last steps. So I would not say that this latest AI victory is really a game-changing success showing that AIs can now beat humans at math. They can be useful for some kinds of problems, but it was obvious decades ago that computers were bound to be useful for things like this.

That's probably what most people would like to know about this latest AI story. Something that maybe should also be interesting about this story, however, and that hasn't gotten attention, is: Why do we care so much about this particular "Navier-Stokes problem"?

Well, it's not clear to me that we do care so much about it. We'd care a huge amount about a way to solve the Navier-Stokes equations in general, but that is not at all what this is. The Clay Institute was founded by some rich guy, and as a sort of publicity stunt for math it defined a set of challenge problems with million-dollar prizes attached. They're all pretty abstract mathematical problems. The one that I believe is considered most important by mathematicians—the Riemann Conjecture—is so arcane that it takes quite a while to even explain what it is, and then it takes much longer to explain why the heck that's important. So for all I know the Clay Institute's "Navier-Stokes problem" was a fairly arbitrary choice by an institute looking to offer prize money as a publicity stunt.

On the other hand, there is something quite interesting in the Clay Institute's specific "Navier-Stokes problem" from the point of view of physics. It's not that the weird kind of fluid flow that has now been shown to be allowed by the NS equations is something that might ever come up in Earth's weather, or anything like that. The interesting physics issue is about whether the Navier-Stokes equations are a fully self-consistent theory, or whether it could ever happen that the mask would slip and the NS equations would betray the fact that they aren't the whole truth.

In fact we know that the Navier-Stokes equations aren't the whole truth. They represent fluids as continuous stuff, smoothly filling up space, when in fact all liquids and gases are collections of zillions of tiny individual molecules, moving around. The Navier-Stokes equations were developed in the early 19th century, at a time when the particle theory of matter was not yet firmly established. The Navier-Stokes equations are a dumbed-down theory that sweeps all the complexity of molecular motion under the rug and talks about larger-scale things, forests rather than trees, as it were. Maybe it would be better spin to call the NS equations a high-level theory, instead of saying "dumbed-down".

There's nothing wrong with that kind of dumbing-down, actually. High-level theories are good, and actually science is full of them. We currently don't believe that we know the ultimate ground-level theory. The most fundamental theory we have—quantum field theory—is believed to be a high-level theory that sweeps lots of things under the rug, and we just don't know what those things are. And every scientific theory we have, short of quantum field theory, is some kind of dumbed-down, high-level caricature of quantum field theory. Science is all about finding theories like the Navier-Stokes equations, that simplify the real story enormously but that seem to be self-consistent enough that one can get away in practice with pretending that they are the whole story.

So it's an interesting question just how reliably self-consistent a high-level theory actually is. Could it ever happen that the theory would break down, and say something that wasn't just simplified but was utterly wrong, so that we would be forced to patch the bug by resorting to the more detailed theory that underlies the high-level theory? In fact we have lots of examples of that kind of scenario. They're always interesting, though, so finding one more scenario like that, in the Navier-Stokes equations, is indeed worth attention.

And that's the kind of specific scenario that this Millennium Prize problem addressed. The question that has now been answered with a positive example is whether a fluid that started out flowing smoothly and slowly could ever be predicted, by the Navier-Stokes equations, to develop into a flow where the speed was infinite at some particular point. In reality, of course, molecules are never going to move at infinite speed, so since the Navier-Stokes equations think that this could happen, then we now know that they are not a fully self-consistent high-level theory. In effect the AI has found the tricky question that backs the Navier-Stokes equations into a corner and makes them abandon their smooth-talking alibi and blurt out the truth.

We already knew that the NS equations are only a high-level theory. We already know pretty well how to patch them with a more detailed theory of molecular motion. And even if we were looking for hints toward better understanding of particle motion, this Millennium Prize problem wouldn't be any help. So this is no breakthrough in physics.

It's just a nice addition to our existing museum of scenarios in which high-level theories break down and need patching with more detailed theories. And so this latest news is an opportunity to draw public attention to that museum. That's why I've taken the time to write this. It's not that I feel any special calling to preach about levels of scientific description to this board in particular! It's just that I do care about this issue, and want to learn how to communicate it better, so I was going to try to write a little essay about it anyway, just for practice, prompted by this latest AI claim. I figured it couldn't hurt to post the result here, since some people here might be interested. It's better than making a lecture about it at a dinner party, because here people can just click away.

I don't think it's widely enough appreciated how pervasive the concept of higher- and lower-level theories is in science. It means that science has a sophisticated ontology that makes a lot of philosophical discussions about what it means to be real seem bizarrely naïve. I think that the concept of higher and lower levels of description should be more widely known and appreciated. A lot of people kind of recognise the idea, but I don't think many people realise how central it is in natural science—or that it offers a natural way to reconcile science with our individual human concerns.
According to John 3:8, Jesus wrote:The wind blows where it wants to, and you hear its sound, but don't know where it comes from and where it is going. So is everyone who is born of the Spirit.
In the original Greek text, the words for "wind" and "spirit" are the same Greek word. I doubt Jesus was talking about the Navier-Stokes equations, but in the spirit of recognising a good idea wherever we find it, I think we can still appreciate the ancient analogy between wind and spirit.

Whether I feel happy or sad, interested or apathetic, is a high-level description of my brain, like the high-level Navier-Stokes description of zillions of zooming molecules in terms of a wind speed, only more so. Psychology isn't as precisely formulated as the Navier-Stokes equations, because brains are a lot more complex than fluids.

Joy and apathy aren't illusions, though, any more than wind is an illusion. They're real in the way wind is real. They're also simplified versions of the full story, the way that wind is a simplified picture of microscopic molecular blizzards. Sometimes the high-level theory breaks down, and what we need is not talk therapy about thoughts and feelings, but medication of some kind. Some of our feelings may be driven by animal instincts that have their own ancient evolutionary agenda. Some of our decisions may be constrained by the mechanics of how our brains work. That's okay. It doesn't mean our feelings and thoughts are not real.

Re: "The" Navier-Stokes Problem

Posted: Sun Sep 13, 2026 3:12 pm
by Gadianton
I'm stuck on why this is even something that needs to be solved. Clearly, I don't understand the problem. (I'm not looking it up, just going by what you say.) I think I understand the basic idea of an idealized mathematical description. I believe I ran the calculation pv=nrt more than any other single equation in my life (that's not saying much), but it describes "ideal" gasses, not anything in reality. Are the problems with N-S, your very complicated equations, the same basic problem as with the simple gas law equation?

It seems obvious that "a fluid that started out flowing smoothly and slowly could ever be predicted, by the Navier-Stokes equations, to develop into a flow where the speed was infinite at some particular point."

Why wouldn't it predict such a thing? why would that be hard? In my mind, wouldn't that be like assuming Galilean motion could keep happening trivially, without constraint, and approach infinity? I can imagine a car going 100kph getting past by another car going 100kph getting past by another -- why is N-S any different in principle than this?

Should we think N-S should cap the speed because relativity says it can't be infinite (then just adjust for that like in special relativity) or because of something more mundane like friction? Would this be like saying: okay, I think ground speed records are harder and harder to break as friction builds up, but I can use N-S to model a car that increases its speed such as to go 100,000 kph with any assumptions about aerodynamics. Like, I can show a school bus going that fast.

Re: "The" Navier-Stokes Problem

Posted: Mon Sep 14, 2026 9:35 pm
by Doctor Scratch
Thanks for this great post, Physics Guy. It seems quite clear to me, and I am about as much of a layperson vis-a-vis this field as anyone.

If you have the time and/or inclination, would you be willing to try to explain the Riemann Conjecture?

Re: "The" Navier-Stokes Problem

Posted: Tue Sep 15, 2026 2:55 pm
by Some Schmo
We could figure out the low level theories or just go with the flow.

Re: "The" Navier-Stokes Problem

Posted: Tue Sep 15, 2026 6:06 pm
by Chap
Physics Guy wrote: ↑
Sat Sep 12, 2026 3:26 pm
I'm stuck on why this is even something that needs to be solved. Clearly, I don't understand the problem. (I'm not looking it up, just going by what you say.) I think I understand the basic idea of an idealized mathematical description. I believe I ran the calculation pv=nrt more than any other single equation in my life (that's not saying much), but it describes "ideal" gasses, not anything in reality. Are the problems with N-S, your very complicated equations, the same basic problem as with the simple gas law equation?
Since Physics Guy is not turning up, let me have a try at answering your question.

The ideal gas equation PV=nRT is about the way that the pressure, volume and temperature (measured from absolute zero) of a sample of a simplified imaginary gas relate to one another. P, V, and T have their obvious meanings, n is the number of moles of gas in the sample, and R is the universal gas constant. If P is in N/m^2, V is in m^3 and T is in Kelvin (like degrees Celsius, but measured from absolute zero), then R is 8.314 J/(mol.K). At high temperatures and low pressures real gases do behave like this, but under other conditions they don't. So this equation is about the state of a stationary sample of gas, just sitting there waiting to be measured.

The Navier-Stokes equations are much more complex, and deal with quite different things They are an attempt to make a rational analysis of what is happening while a sample of a fluid (liquid or gas) is in motion through changing conditions - e,g, a gas passing from a wide pipe into a narrow nozzle. Pressure comes into it, but so do acceleration, density and viscosity, and external forces such as gravity and the electromagnetic forces present when we are dealing with a flow of plasma. They are not just a different ball game, but a different kind of game altogether.

Re: "The" Navier-Stokes Problem

Posted: Sun Sep 20, 2026 9:18 am
by Physics Guy
Thanks for stepping in, Chap. Yeah, the ideal gas law is a different kind of thing from the Navier-Stokes equations, in several ways.

First of all, it's less fundamental and general, and more of a special case. The ideal gas law is an example of what is called "an equation of state". It characterises a particular category of substances. The ideal gas law was determined empirically before we were sure that matter was made up of particles, but in terms of particles, an ideal gas is a bunch of separate particles flying around and only interacting with each other significantly when they actually run into each other. They're spread out in space sparsely enough that at any given instant only a really tiny fraction of all the particles are actually in the process of colliding; the rest are all just flying straight at some speed. Any gas of that kind, regardless of exactly what the particles are, will be an ideal gas, with pressure directly proportional to both temperature and density.

The ideal gas law is one of the possible rules that could be slotted into the Navier-Stokes equations, to specify what the pressure in the fluid should be. The NS equations include a pressure term, but they don't say what it is; they're kind of like an event planner who knows that the schedule includes a noon meal, but the menu is up to the caterer. The ideal gas law is like the caterer letting you know that you're getting the tofu sliders.

Secondly the ideal gas law is a constraint, prescribing a relationship (in this case, between pressure, temperature, and density) that has to be true all the time, everywhere. The Navier-Stokes equations, on the other hand, are a set of differential equations. They say how things can change, over time and from place to place. The gas law is about things at the same place and time, but the NS equations are about relationships between things at different places and times. One might think that the constraints on things at the same place and time were the more basic laws, with the laws about change coming after, but it's really the other way around. The basic laws of nature, as far as we know, are about how things can change. In a lot of practically important scenarios, however, the way things change turns out to produce a sort of conspiracy in which some things tend to stay constant, in kind of the way that classical economics predicts that free markets will tend to maintain stable prices, except that in physics this actually works. So if you start with the basic laws about how particles move, you can deduce the ideal gas law as a nearly inevitable feature of low-density gases.

Why is this new conclusion, that the NS equations can allow infinite flow speed to develop, even surprising? I think I forgot to mention that the infinite flow speed develops, at least at one point, after only a finite amount of time. So somehow a little bit of the fluid has undergone infinite acceleration, which seems as though it should require infinite force, but the problem stipulates that the external forces are all finite. So somehow the NS equations are allowing a tiny patch of fluid to sort of bootstrap its way to infinite speed, from forces within the fluid itself. Even apart from relativity, this is crazy. The equations weren't supposed to permit things like that. It's kind of an exploit, an escape from the sandbox.

It's not a total shock because the NS equations were only ever an approximation, and it was never obvious that glitches like this couldn't happen. They only seem to happen in pretty far-fetched scenarios, so the NS equations are still fine to use in all their many important practical applications.