Why 3D Random Walks Are Fundamentally Different

追加: 2026-05-09

[00:00:00]Imagine you're standing on a number line at zero. Every second you flip a coin, heads you step right, tails you step left. Will you ever come home? The answer is yes. And not just sometimes.

[00:00:14]With probability one, no matter how far you wander, you'll return to zero eventually. Maybe in a million steps, but you'll come back. Now do the same thing on a 2D grid with four directions.

[00:00:27]You roll a four-sided die. Up, down, left, right. Same question. Will you come home? Almost everyone says no. Two dimensions feel bigger. More room to escape. They're wrong. You still come home with probability one, but step up to three dimensions and something snaps.

[00:00:48]In three dimensions with six directions, the answer flips. A random walker in 3D space has a real chance of never coming back. About 66%.

[00:00:59]The drunk bird gets lost forever. That's a quote from George Paulia who proved this in 1921.

[00:01:06]A drunk man will find his way home, but a drunk bird may get lost forever. The shift happens between dimension 2 and dimension 3. Two is the borderline case, a close call, just barely recurrent.

[00:01:20]Three is where the world breaks. And what's stranger, this isn't an artifact of how we set up the lattice. The same dichotomy holds for Brownian motion, for diffusion, for any natural notion of random wandering in space. Two is the last dimension where you have to come home. This video is about why. Why does the universe care about the difference between flat and three? Okay, let's start in one dimension to see the simplest case. You're at zero. You take a step left or right with equal probability. After one step, you're at plus or minus one. After two steps, you're at minus2, 0, or plus two. After many steps, you've drifted somewhere.

[00:02:03]The thing is, the walker spends most of its time near zero. Not because anything is pulling it there, because most paths cancel. A walk of length 2 n returns to zero exactly when you've taken n lefts and n rights. The probability of that is binomial coefficients divided by powers of two. Using Sterling's approximation, this works out to roughly one over the square root of n. The walker comes back to zero often with that frequency. And if you ask across infinitely many steps, will it return at least once? Sum that probability over all even n. You get an infinite number. So returns are certain.

[00:02:44]Now jump up a dimension. You're standing on a grid. Each step picks one of four neighbors at random. The grid is the standard integer lattice in two dimensions. Looking at it, intuition says more space, easier to escape. After all, a 1D walker is locked on a line.

[00:03:02]Its options are limited. A 2D walker has a whole plane to wander into. But the math is going to surprise you. The 2D walker also returns to the origin with probability one. always, no matter how complicated its path looks. And there's a beautiful trick that makes it obvious.

[00:03:20]Once you see it, we'll get there. First, let's actually watch one of these walks for a while. Here's a single 2D random walk traced out for several thousand steps. Notice how dense the path becomes near the origin. It keeps coming back.

[00:03:36]It explores. It crosses itself constantly. If we let this run long enough, it fills a region. A theorem by Palia tells us something stronger. Not only does the walker return to zero with certainty, it visits every single lattice point, every integer coordinate with probability one given enough time.

[00:03:58]It eventually touches every node of the grid. So 2D is recurrent in the strongest possible sense. You're not getting away from any specific point.

[00:04:08]And yet almost every visualization of a 2D walk looks like it's drifting somewhere. That's because at any finite time, the walker is typically far from origin. Returns happen. They're guaranteed, but they take longer and longer. The expected return time is infinite. Even though the return itself is certain, that's a strange combination, and we'll come back to what it means. Now lift the walk into three dimensions, six directions, up, down, left, right, forward, back. Pick one uniformly at random. Take a step. Same setup, same kind of randomness, just one more dimension, and suddenly you might not come home. Look, for a 3D random walker starting at the origin, the probability of ever returning is about 0.34.

[00:05:01]twothirds of the time it just leaves.

[00:05:04]This is so different from 1D and 2D that it's hard to believe. Same rules, same fairness. The walker doesn't know what dimension it's in, but the geometry of three-dimensional space lets it escape.

[00:05:17]Let's actually watch. This is a 3D random walker traced over a few thousand steps. The camera rotates around the path so you can see how it lives in the volume. At first, it looks like the 2D version. Close to origin. Lots of folds crisscrossing. Then slowly it commits to a direction. It drifts. There's no force pulling it. Each step is independent and balanced. But the geometry of three-dimensional space gives it more places to be. And once a path has wandered far enough, the chance of coincidence, the chance of returning drops fast. Watch the path keep going as the camera circles around. You can see the walker getting further and further from the origin point. It's not coming back. And if you ran a thousand independent walks side by side, about 66% of them would do exactly this.

[00:06:11]Drift, never return. The number we mentioned, 0.34, is the probability that a 3D walk ever returns to its starting point. That number was first computed by an English mathematician named GN Watson in 1939.

[00:06:28]It's the answer to a triple integral that nobody can simplify. We'll see it shortly. For now, the key thing, this picture, this drift, this slow commitment to escape. It's not an illusion. Most 3D walks really do leave forever. Let's just put the number on screen. The probability a 3D walker returns to origin is 1 - 1 / u of3 where u of 3 is Watson's integral. That's a triple integral over a cube with an integrant involving the cosiness of three coordinates. Numerically u of 3 is about 1.516 and 1 - 1 over that is about 0.34.

[00:07:10]So, a 3D walk returns about 34% of the time or escapes about 66% of the time.

[00:07:18]It's a specific number with no closed form. You can't write it as a simple fraction or an algebraic expression. And yet, it falls right out of the random walk on the integer lattice in three dimensions. A specific physical setup gives you a specific transcendental number. This kind of thing happens a lot in math. Okay. So why is dimension two the cutoff and not three or four? The answer comes from one elegant test due to Paulia himself. A random walk is recurrent, guaranteed to come home if and only if a certain sum diverges. The sum is over n of the probability of being back at the origin after n steps.

[00:08:01]If that infinite sum adds up to infinity, the walk is recurrent. If the sum converges to a finite number, the walk is transient. The whole question is decided by one infinite series.

[00:08:13]Convergence of a sum. The intuition is simple. The sum is the expected number of visits to the origin across all time.

[00:08:22]If that expected number is infinite, returns happen forever. If it's finite, the walker only visits a finite number of times before leaving for good. And the terms of that sum decay at different rates depending on dimension. In 1D the terms decay like 1 over the square roo<unk> of n. In 2D like 1 / n. In 3D like 1 / n^ 3s. Two of these sums diverge. The third converges. The dimension where the cutoff happens is hidden in the exponent. The 1D walk returns to origin after 2n steps with probability given by C of 2n choose n / 4 to the n. For large n, sterling's formula tells us this is approximately 1 / the square<unk> of pi n. So the term decays like 1 /<unk> n. The sum of 1 /<unk> n from n = 1 to infinity diverges. You probably already knew this. The harmonic series 1 plus a half plus a third and so on diverges. The sum of 1 /<unk> n diverges even faster. It's a bigger sum. So the 1D walk is recurrent. The series sums to infinity.

[00:09:37]And by criterion, the walker comes home.

[00:09:40]Now here's where it gets beautiful. The 2D case has a clever decoupling argument that makes it work, too. Here's the trick. A 2D walker steps in one of four directions, right, left, up, down. Look at the same walk in rotated coordinates.

[00:09:57]Define U as X + Y and V as X minus Y. A right step changes X by + 1, Y by 0, so U goes up by 1, v goes up by 1. A left step changes X by minus1, Y by 0. So, u goes down by one, V goes down by one. An up step changes X by zero, Y by + one.

[00:10:22]So, U goes up by one, V goes down by one. And a down step does the opposite.

[00:10:28]U goes down, V goes up. Watch what just happened. In the new coordinates, every step changes U by plus or minus1 and V by plus or minus1. And you can check U and V change independently each with probability 1/2. The 2D walk became two independent 1D walks. Now the probability of being at origin after 2n steps is just the probability that the U walk returns times the probability that the VWalk returns. Each is around 1 /<unk> N. So the product is around 1 /n.

[00:11:04]The sum of 1 /n diverges barely. That's the harmonic series. So the 2D walk is recurrent just like the 1D walk in three dimensions. No such trick exists. The sixstep directions don't decouple into three independent 1D walks. You can still count paths. You can still apply local central limit theorems. And it turns out the probability of being at the origin after 2n steps decays like 1 / n to the three halves. That's a faster decay. The sum of 1 / n^ the three halves converges. It's a finite number.

[00:11:42]By Polia's criterion, the walk is transient. The walker has positive probability of never returning. The exponent crossed the line between divergence and convergence. Dimension two is the borderline. Dimension three is over the edge. This dimensional dichotomy shows up everywhere. Brownian motion, the continuous version of random walk, has the same story. 1D and 2D Brownian motion is recurrent. 3D Brownian motion is transient. It drifts away with probability one. Electrical networks too. You can map a random walk on a lattice to current flowing through a network of 1 ohm resistors. The walk is recurrent if and only if the resistance from origin to infinity is infinite. In 2D, that resistance grows logarithmically. In 3D, it stays finite.

[00:12:35]A photon scattering through a thin layer of gas behaves differently than one scattering through a volume. A polymer chain folded onto a flat surface follows different statistics than one folded in space. Even physical search and diffusion limited reactions split at this boundary. Two and three are different worlds. And every time you encounter a phenomenon involving diffusion or scattering or random search in physical space, dimension two and dimension 3 behave differently always.

[00:13:08]So there's a real concrete sense in which two-dimensional space is fundamentally different from three-dimensional space. A drunk man on a 2D grid will always find his way home.

[00:13:20]A drunk bird in 3D space, given enough time, will probably leave forever. The proof comes down to a single sum diverging or converging and the boundary lies exactly between dimensions 2 and three. Palia found this in 1921. The number 0.34 took until 1939. It's a small fact with a long shadow that touches everything from physics to network theory.