Krylov-Bogoliubov theorem

In any system that evolves over time—from a planet orbiting a star to the fluctuations of a financial market—we often seek a sense of long-term predictability. While individual future states may be chaotic or random, is there a stable statistical pattern, a distribution of states that remains unchanged by the system's dynamics? This concept of statistical equilibrium is captured by the mathematical idea of an ​​invariant measure​​. The critical question that arises is refreshingly direct: for a given dynamical system, how can we be certain that such a stable state even exists? Without this guarantee, the search for long-term statistical laws could be a fruitless endeavor.

This article explores the profound answer provided by the ​​Krylov-Bogoliubov theorem​​ and its powerful extensions. In the chapter "Principles and Mechanisms," we will uncover the elegant logic behind the theorem, revealing why compactness is key to guaranteeing existence and how this idea is adapted for systems that roam infinite spaces. Subsequently, in "Applications and Interdisciplinary Connections," we will witness the theorem in action, showing how it brings order to the apparent randomness of chaotic systems, describes the equilibrium of physical processes, and even tames the infinite complexity of turbulent fluids. Let's begin by dissecting the core argument and the mathematical magic that ensures order can be found within change.

Imagine you release a single drop of ink into a swirling glass of water. At first, its path is a chaotic, unpredictable dance. But wait long enough, and the ink spreads out, eventually reaching a state of equilibrium—a uniform, pale gray distribution throughout the water. This final, stable state is a distribution that no longer changes, even as the water molecules continue their ceaseless motion. In the world of mathematics and physics, we call such a state of statistical equilibrium an ​​invariant measure​​. It is a snapshot of a system's long-term behavior, a probability distribution that remains unchanged by the system's evolution.

The fundamental question, which the Krylov-Bogoliubov theorem answers, is profound in its simplicity: for a given dynamical system, can we be certain that such a stable state, such an invariant measure, even exists?

Let's begin with the most natural idea imaginable. If you want to know where a system tends to spend its time, just watch it for a very long time and keep a running tally. Let’s picture a point $z_0$ on a circular disk, and a rule, a continuous function $T$, that tells us where the point jumps to in the next instant. The path, or ​​orbit​​, of the point is the sequence $z_0, z_1=T(z_0), z_2=T(z_1), \dots$.

To describe where the point has been, we can construct what is called an ​​empirical measure​​. After $N$ steps, this measure, let's call it $\mu_N$, is simply the average of the locations the point has visited:

Here, $\delta_z$ is a ​​Dirac measure​​, which you can think of as a "point mass" located entirely at the point $z$. So, $\mu_N$ is a collection of point masses, each with a tiny weight of $1/N$, scattered across the orbit.

As we let $N$ grow larger and larger, we expect this cloud of points to give us a picture of the system's long-term habits. If this sequence of empirical measures converges to some limiting measure $\mu$, common sense suggests this limit ought to be invariant. Why? Let's consider the difference in the average value of some observable quantity, say a function $f$, before and after one step of the dynamics. This difference can be written as:

The notation $\int f \, d\mu_N$ just means "the average value of $f$ over the first $N$ points of the orbit." A beautiful thing happens when we expand this expression. The sum becomes a ​​telescoping series​​: all the intermediate terms cancel out, leaving just the endpoints.

As $N \to \infty$, the denominator grows without bound. If our function $f$ is bounded (which is true for any continuous function on a compact space), the numerator remains finite. So, the difference $\Delta_N(f)$ must go to zero! This tells us that in the limit, the measure $\mu$ must satisfy $\int (f \circ T) \, d\mu = \int f \, d\mu$. It is invariant! This simple, elegant argument is the intuitive heart of the matter.

The argument above relies on a crucial assumption: that the sequence of measures $(\mu_N)$ actually converges to something. Why should it? In the infinite-dimensional world of measures, this is not a trivial question. The answer lies in one of the most powerful concepts in mathematics: ​​compactness​​.

A compact space is, loosely speaking, a space that is "contained" and "complete." The closed unit disk is compact; the entire infinite plane is not. The magic of compactness is that any infinite sequence of points within a compact space must have a "cluster point"—a point around which infinitely many members of the sequence gather.

Now for the brilliant leap, formalized by the Banach-Alaoglu theorem. The space of all possible probability measures on a compact space $X$, denoted $\mathcal{P}(X)$, is itself a compact space in a certain sense (specifically, in the weak-* topology). Think about it: if our point is confined to a compact disk, then any statistical distribution of its location is also confined in a meaningful way.

Because our entire sequence of empirical measures $(\mu_N)$ lives inside this compact world of $\mathcal{P}(X)$, it is guaranteed to have at least one cluster point. And as we've just seen, any such cluster point must be an invariant measure. This is the essence of the ​​Krylov-Bogoliubov theorem​​: for any continuous map on a non-empty compact space, an invariant probability measure is guaranteed to exist.

This is wonderful, but what if our system doesn't live on a nice, tidy compact space? What if it's a particle moving in the vastness of $\mathbb{R}^d$? Now we have a serious problem. The particle might just wander off to infinity, never to return.

Consider a process described by the stochastic differential equation (SDE) $dX_t = X_t\,dt + dW_t$. The drift term $X_t\,dt$ actively pushes the particle away from the origin. The particle's position will tend to grow exponentially. If we average its position over a long time $T$, the average will just keep growing as $T$ increases. The probability "mass" of our empirical measure escapes to infinity. In this case, the time-averaging procedure fails to produce a probability measure; it converges vaguely to the "zero measure," which tells us nothing.

To ensure existence of an invariant measure on a non-compact space, we need a guarantee that the system is, in some sense, self-confining. This property is called ​​tightness​​. A family of measures is tight if, for any small probability $\epsilon$, you can find a single, fixed, large-enough box (a compact set) that contains the system with probability at least $1-\epsilon$, uniformly for all the measures in the family. Tightness is the promise that probability mass does not leak away to infinity.

A beautiful example highlights this subtlety. Consider a particle whose drift pulls it back toward the origin, governed by $dX_t = -\beta\,\dfrac{X_t}{1+X_t^2}\,dt + dW_t$. It turns out that everything depends on the strength of this pull, $\beta$. If $\beta > 1/2$, the pull is strong enough to keep the particle contained, the empirical measures are tight, and an invariant measure exists. But if $\beta \le 1/2$, the pull is too weak. The particle can still make long, "heavy-tailed" excursions from which it takes too long to return. The system is recurrent (it will always come back eventually), but it's not "positive recurrent" (the average return time is infinite). The probability density becomes non-normalizable, and tightness is lost.

How can we prove a system is tight? We need a tool to certify that it's self-confining. This tool is the ​​Lyapunov function​​. Imagine a function $V(x)$ that measures the "energy" or "distance from home" of the system. For example, $V(x) = |x|^2$ is a simple choice. We require this function to be ​​proper​​ (or coercive), meaning $V(x) \to \infty$ as $|x| \to \infty$.

The crucial insight is to look at the expected rate of change of this energy, which is given by the system's generator $\mathcal{L}$ acting on $V$. The ​​Foster-Lyapunov drift condition​​ states that there must be a "drift towards the center". In its simplest form, it says that whenever the system is outside some large compact set $K$, its energy tends to decrease:

where $c$ is some positive constant. This inequality is like a cosmic leash. The farther the particle strays, the more strongly it's pulled back. This condition guarantees that the expected time to return to the "home" region $K$ from any starting point $x$ is finite, and is in fact bounded by a multiple of its initial "energy" $V(x)$. This return property is called positive recurrence, and it is the key that unlocks tightness.

Revisiting our boundary-case example, for $dX_t = -\beta\,\dfrac{X_t}{1+X_t^2}\,dt + dW_t$ with $\beta > 1/2$, we can choose $V(x)=x^2$ and show that the Foster-Lyapunov condition holds. The weak confining force is just strong enough to ensure tightness and the existence of an invariant measure.

We can now assemble the full picture for a general stochastic process. To prove the existence of an invariant measure, we need two key ingredients:

1. ​​Tightness:​​ The family of time-averaged measures $\{\mu_T^x\}$ must be tight. This is the recurrence condition, preventing escape to infinity, often established by a Foster-Lyapunov condition. By ​​Prokhorov's theorem​​, tightness guarantees our sequence of measures has a limit point that is a true probability measure.
2. ​​The Feller Property:​​ The dynamics must be sufficiently smooth, in the sense that they map continuous functions to continuous functions. This technical requirement is what allows us to validly take the limit in the invariance equation and prove that the limiting measure is, in fact, invariant.

Together, these conditions guarantee the existence of at least one invariant measure. But is there only one? For our ink in water, we expect a single, unique state of equilibrium. To guarantee uniqueness, we need an extra ingredient: ​​irreducibility​​. This property means the system can, with positive probability, travel from any open region to any other open region. There are no inescapable pockets or separate universes. When the system is both recurrent (ensuring existence) and irreducible, it typically possesses a unique invariant measure.

This leads to the beautiful concept of ​​ergodicity​​. For a system with a unique invariant measure $\pi$, the time average along almost every trajectory converges to the same space average given by $\pi$. Your single, long-term observation of the system will reveal its one and only true statistical nature.

Finally, let us step back and admire the entire landscape of stable states. The Krylov-Bogoliubov theorem tells us that the set of all invariant measures, $\mathcal{M}_T(X)$, is not just non-empty. For a system on a compact space, this set is also ​​convex​​ (if you have two stable states, any weighted average of them is also a stable state) and ​​compact​​. This means the collection of all possible equilibria is a well-behaved, solid geometric object living in the abstract space of measures. The search for stability does not lead us to a scattered group of isolated points, but to a unified, structured whole.

In our last discussion, we uncovered a wonderfully reassuring piece of mathematical truth: the Krylov-Bogoliubov theorem. It tells us that if we have a system whose evolution is continuous and confined to a bounded, closed (or "compact") playground, it cannot wander aimlessly forever. Sooner or later, it must settle into a pattern of behavior, a kind of statistical rhythm described by what we call an invariant measure. This is a profound promise of order beneath the surface of change.

But, you might rightly ask, what is this promise good for? The real world isn't always a neat, compact playground. A particle can wander through all of space; a turbulent fluid seems to possess an infinite number of ways to twist and turn. Where, in this vast, untidy universe, do we find these invariant measures, and what secrets can they tell us? This is our next adventure: to journey from the theorem's ideal world into the realms of physics, chaos, and even the infinite, and see how this single, elegant idea brings clarity to them all.

The theorem’s requirement of a compact space seems, at first, to be a serious limitation. But nature is clever. Often, even when a system has an infinite space to explore, there is a "tether"—a restoring force that pulls it back towards a home base, effectively confining its motion.

Consider one of the most fundamental models in all of physics: the Ornstein-Uhlenbeck process. Imagine a tiny particle suspended in a fluid, being constantly kicked about by the random jiggling of water molecules—a path we call Brownian motion. Now, let's attach a tiny, perfect spring from this particle to a fixed point. If the particle strays too far, the spring pulls it back. The particle is free to move anywhere in space, yet it is not truly free. This is the essence of the Ornstein-Uhlenbeck process. It describes everything from the velocity of that dust mote in a sunbeam to the fluctuations of interest rates in financial markets, which tend to revert to a long-term average.

How does this "tethering" help us? Here we use a beautifully intuitive idea, given a rigorous footing by the great mathematician Aleksandr Lyapunov. We can define a quantity, a sort of "energy" or "unhappiness function," that measures how far the system is from its comfortable home. For our particle, let's use the simple function $V(x) = 1 + \|x\|^2$, where $\|x\|$ is its distance from the origin. Now, we ask: how does the system's evolution, on average, change this value? The mathematical machinery of stochastic calculus shows us something remarkable. The combination of the spring's pull and the viscous drag from the fluid conspires to create a "drift" in this energy, governed by an inequality like $\mathcal{L}V(x) \le -\alpha V(x) + \beta$.

Don't be intimidated by the symbols. All this says is that the further the particle is from home (the larger $V(x)$ becomes), the stronger the negative push on its energy, pulling it back. This guarantees the particle cannot escape to infinity. It ensures that the family of probability distributions describing its position is "tight"—it doesn't spread out indefinitely. This tightness is the key that unlocks the door for the Krylov-Bogoliubov machinery in a non-compact world. It guarantees that a statistical steady state, an invariant measure, must exist.

And what is the shape of this steady state? When we solve for it, we find something extraordinary. The tug-of-war between the random outward kicks and the deterministic inward pull is perfectly balanced to produce a Gaussian distribution—the iconic bell curve. The final probability density $p_\infty(x)$ for the particle's position is given by

where the parameters $\alpha$ and $\beta$ are related to the strength of the spring and the intensity of the random kicks. This is not just a mathematical formula; it is a portrait of equilibrium, a universal pattern that emerges whenever a restoring force battles against random noise.

Some systems, however, don't settle into a placid bell-curve. They dance. In the realm of chaotic dynamics, a system's state may leap unpredictably for all time, never repeating, yet still be confined to a beautifully intricate, bounded region known as a ​​strange attractor​​. This attractor is the compact playground where our theorem comes to life.

A classic example is the logistic map, a deceptively simple equation $x_{n+1} = 4x_n(1-x_n)$ that was a cornerstone in the discovery of chaos. The value of $x$ jumps around the interval $[0, 1]$ in a way that seems utterly random. But it is not. The Krylov-Bogoliubov theorem guarantees an invariant measure exists, and for this system, we can find it explicitly. It has a density known as the arcsine distribution, $\rho(x) = 1/(\pi\sqrt{x(1-x)})$. This density is a statistical map of the chaos. It tells us that the system spends most of its time near the endpoints (0 and 1) and is least likely to be found in the middle. With this map in hand, we can predict the long-term average of any property of the system, transforming a chaotic dance into a predictable statistical ballet.

The power of this idea becomes even more apparent with more complex systems like the Hénon map, which generates a stunning fractal attractor in the plane. This attractor is the system's true home, a compact set where the dynamics unfold. The corresponding invariant measure, often called the "physical measure," tells us the probability of finding the system in any given region of the attractor. Its very invariance becomes a tool of immense power. Just as conservation of energy helps us solve complex mechanics problems, the invariance of the measure—the fact that the statistical distribution looks the same after one step of the evolution—allows us to derive relationships between the statistical moments of the system. Incredibly, this can allow us to calculate properties like the attractor's center of mass, using algebraic relations born from the principle of invariance itself, a feat that would be impossible by simply watching the chaotic trajectory unfold.

What happens when we move from a few variables to a system with infinite degrees of freedom—a field, a vibrating string, or a fluid in motion? This is where the theorem faces its greatest challenge and achieves its most stunning successes.

Let's begin by considering a stochastic process, like a swarm of diffusing particles, confined to a box. The box is a bounded, compact domain. If the walls of the box are "reflecting," meaning any particle that hits a wall is simply bounced back in, then no particles are ever lost. The total probability of finding a particle inside the box remains constant, fixed at one. In this closed system, the conditions of the Krylov-Bogoliubov theorem are met, and an invariant probability measure—a steady-state distribution of particles—is guaranteed to exist. But if the walls are "absorbing," acting like open windows through which particles are lost forever, then the total probability inside the box dwindles to zero. The system is not closed, and no invariant probability measure can be found. This simple example gives us a profound physical intuition: invariant measures correspond to the steady states of closed systems that conserve their "stuff," be it particles or probability.

The true leap comes when we consider a fluid. A complete description of a turbulent flow, even in a teacup, requires an infinite number of variables to specify the velocity at every single point. A bounded set of fluid configurations in this infinite-dimensional space is not, in general, compact. For a time, this seemed an insurmountable barrier.

The solution is a masterpiece of mathematical physics, bringing together all the ideas we have discussed. Consider the stochastic Navier-Stokes equations, the laws governing a fluid that is being randomly stirred. We can write down an energy balance for the fluid. The random stirring pumps energy in, while the fluid's own internal friction, its viscosity, dissipates it. In two dimensions, this dissipation is remarkably effective. It preferentially damps out the small-scale, high-frequency whorls and eddies. This has a miraculous consequence: it ensures that while the fluid's energy is bounded, the "smoothness" of the fluid flow is also bounded on average.

Here is the magic key: the Rellich-Kondrachov theorem, a deep result from analysis, tells us that a collection of fluid configurations that is bounded in "smoothness" is a compact set when viewed in the space of ordinary fluid configurations! It is the infinite-dimensional echo of our Lyapunov function. The viscous forces act as a tether on the high-frequency motions, confining the dynamics to what is effectively a compact region of the state space. The Krylov-Bogoliubov logic takes over, and we can prove that even a randomly-driven, turbulent fluid will settle into a unique statistical equilibrium, an invariant measure that governs its climatic properties.

This leads to a final, breathtaking insight. What if we only stir the fluid in a very specific, limited way—say, by kicking just a handful of its largest eddies? Common sense might suggest that the randomness would remain confined to those large scales. But Nature is more unified. The fluid's own internal dynamics—the nonlinear way that large eddies cascade down to create smaller and smaller ones—can grab the randomness from the few excited modes and spread it throughout the entire system. If the forcing, however limited, satisfies a special "saturating" condition related to this nonlinear mixing, the system becomes fully irreducible. It will explore every possible state it can reach, eventually settling down to a single, unique invariant measure. The interplay between the limited random forcing and the rich deterministic nonlinearity gives rise to a globally unique statistical state.

From a particle on a spring to the chaos of a strange attractor and the roiling of a turbulent sea, the principle of the invariant measure stands as a testament to the emergence of statistical order. The Krylov-Bogoliubov theorem is not just an abstract statement; it is a lens through which we can see the universal rhythms that govern the long-term behavior of our complex and wonderful world.