Stand at a corner of an infinite grid and start to wander — at every tick, one step in a direction chosen by a fair coin, no memory of where you have been. Question with a yes-or-no answer and a famous one: are you certain to come back to the corner you started from?
On a line, yes. On a flat plane, also yes — return is certain, probability exactly one, though you may have to wait a very long time. Add a third dimension — let the walker fly, up and down as well as the four compass steps — and the certainty breaks. The flying walker comes home only about 34% of the time. Two times in three it drifts off and never returns. George Pólya proved this in 1921: the simple symmetric random walk is recurrent in one and two dimensions and transient in three and above. The plane keeps you; volume lets you escape. The difference is one dimension too many.
Watch one walker
In 1-D and 2-D returns keep coming almost surely, forever: teal dots mark them in 1-D, flashes in 2-D. In 3-D most walkers never return at all. Among those that do, there is almost surely a last return, after which the flashes stop. One walk can't prove the theorem; a crowd below can illustrate the difference.
Run a crowd of them
Certainty is a statement about the whole ensemble, so let many walkers run and count what fraction touches home again within a fixed horizon. A finite horizon misses some later returns, so the probability of returning by then is below the eventual-return probability. The sampled fraction also fluctuates; in 3-D it can land above or below the true limit.
Gold line = the true return probability. As you extend the horizon, the finite-time return probabilities approach 1 in 1-D and 2-D (slowly in 2-D), and about 0.34054 in 3-D. The sampled bars fluctuate around those finite-time probabilities. The limiting missing fraction in 3-D is about 65.95%: gone for good.
Why two is the borderline
The proof is a single sum. Let p₂ₙ be the probability the walker is back at the origin after 2n steps (only even times can return). The walk is recurrent exactly when the expected number of returns is infinite — when Σ p₂ₙ diverges. A local central-limit estimate pins the size of each term:
So the terms shrink like n−d/2. In 1-D that is n−1/2 and the sum diverges; in 2-D it is n−1 — the harmonic series, which still just barely diverges; in 3-D it is n−3/2, and that sum converges. Two dimensions is the knife-edge: the walker comes home with probability one, but the expected wait is infinite. Three dimensions tips over the edge, and the total number of returns becomes finite.
The exact number
When the sum converges, it converges to something nameable. Its value u(3) = Σ p₂ₙ is the expected number of visits to the origin (counting the start), and the return probability is p = 1 − 1/u. In 1921 this was a limit nobody could evaluate; in 1939 G. N. Watson evaluated the triple integral that u(3) hides inside using elliptic integrals. Glasser and Zucker (1977) later expressed the result in Gamma functions:
So the drunk bird's odds of ever getting home are about 34.05% — a number with a Gamma-function pedigree, not a round one. The same machinery gives the odds in higher dimensions, and they keep falling: more room to get lost in means less chance of stumbling back.
| dimension | return prob p(d) | verdict | chance of coming home |
|---|
1-D and 2-D return with certainty (p = 1); from 3-D on, the bird escapes. The displayed d≥4 decimals are numerical values (Finch, Mathematical Constants, §5.9). Gaunt, Nadarajah and Pogány (2024) give exact expressions in Lauricella functions for d≥3, without a further simplification for d≥4. Sources for every figure are in the colophon.
Where the picture bends, honestly
The lattice is a model. The claim here is about simple symmetric walks on the full integer lattice ℤᵈ. Mean zero and finite variance alone do not suffice: a walk confined to a line inside 3-D is still a 1-D walk. Brownian motion has the same dimension threshold for neighborhood recurrence; in 2-D it revisits every neighborhood but almost surely never hits the exact starting point again (Bishop and Peres, p. 229). Heavy-tailed jumps or drift can change the behavior.
The Monte-Carlo bars miss late returns. A walker that hasn't come home by step 4,000 might still return at step 40,000, in 3-D as well as 1-D and 2-D. Extending the horizon raises the finite-time return probability in all three dimensions. Individual samples fluctuate; only the limiting escape probability in 3-D is permanently missing.
The aphorism has a documented attribution. Durrett (2019), p. 289, attributes the joke to Kakutani at a UCLA colloquium, without giving the talk's date. The wording above is a traditional variant: Durrett writes “will eventually find”. His passage does not establish the original wording from a transcript or recording.
Two dimensions is certain but cruel. "Recurrent" means return with probability one — it does not mean soon. In 2-D the expected time to return is infinite (that harmonic-series knife-edge again). You will get home; do not hold your breath.
- 1-D, 2-D, 3-D return terms p₂ₙ by exact BigInt counting, matched against brute-force path enumeration
- the local-CLT size 2(d/4πn)d/2 confirmed against the exact terms (ratio → 1)
- u(3) three ways: series + asymptotic tail ≈ 1.51624; Watson's integral by quadrature ≈ 1.51378; Gamma form ≈ 1.5163860592. The series/integral and integral/Gamma comparisons allow absolute differences up to 0.02
- p(3) = 1 − 1/u(3) = 0.3405373296; higher-d constants p(4)≈0.193, p(5)≈0.135 by lattice quadrature
- Monte-Carlo: 3-D return fraction lands near a third; 1-D far higher — recurrence vs transience seen, not assumed
29/29 checks pass in research/polya-recurrence/verify.mjs — pure Node, no dependencies, run with node research/polya-recurrence/verify.mjs.
One dimension is a tightrope you keep falling back onto. Two is a floor that always, eventually, catches you. Three is open sky — and in open sky, most things that leave do not come back. The whole theorem is in the gap between those last two, and the gap is exactly one coordinate wide.