The orbit that must come back

Start a system, let it run, and wait. Does it ever come back to where it started?

For a system with finitely many states the answer is immediate. There are only so many states, the rule is deterministic, and a sequence drawn from a finite set has to revisit one — more things than boxes, applied to time.

The interesting question is what happens when the states form a continuum. There are infinitely many, no two need ever coincide, and the pigeonhole argument has nothing to grip. Poincaré’s answer, from 1890, is that the conclusion survives anyway, in a slightly weakened form that turns out to be enough.

The statement #

Poincaré recurrence. Take a system whose state space has a finite total volume, evolving by a rule that preserves volume. Take any region, however small. Then almost every point of that region returns to it — and, since the argument can be repeated, returns infinitely often.

Two hypotheses do the work and both are essential.

Finite volume. The system cannot run away. A ball rolling on an infinite plane never returns; a ball in a bowl has nowhere to go.

Volume preservation. As the region is carried around by the rule, its volume stays the same. It may be stretched into a filament of enormous length and microscopic width, but the product does not change.

Given those, the conclusion is forced, and the argument is short enough to give in full.

The argument #

Let $A$ be the region and let $A_n$ be where $A$ has got to after $n$ steps. Each $A_n$ has the same volume $v > 0$, and all of them sit inside a space of total volume $V < \infty$.

Now suppose no point of $A$ ever returns to $A$. Then $A$ and $A_n$ are disjoint for every $n \ge 1$. Applying the rule repeatedly, $A_m$ and $A_n$ are disjoint for every $m \neq n$ — the rule carries a disjoint pair to a disjoint pair.

So the space contains infinitely many disjoint regions each of volume $v$. Their total volume is infinite. But the space has volume $V$, which is finite. Contradiction.

That is the whole proof. It is the pigeonhole principle with volume in place of cardinality: infinitely many things of the same positive size cannot fit into a container of finite size, exactly as infinitely many pigeons cannot fit into finitely many holes one apiece.

The conclusion is weaker than in the finite case, and the weakening is the phrase “almost every”. The finite version gives an exact repeat with a computable bound; this one gives a near return with no bound at all. Some points genuinely never return — the ones sitting exactly on a repelling fixed point, for instance — but the set of them has volume zero, so a point picked at random is not one of them.

What it does not say #

Three readings are tempting and wrong, and each is worth removing.

It does not say the orbit repeats. Returning near the start is not returning to it. The rotation above never revisits a point exactly, and Poincaré’s theorem does not claim it does. What comes back is the region, not the point.

It does not say when. The proof establishes that a return happens and gives no bound whatever on how long it takes. That silence is not a gap in the argument; the time genuinely can be arbitrarily long, and how long is the subject of the next section.

It does not apply to dissipative systems. A pendulum with friction has volume-contracting dynamics — regions shrink, and the argument fails at the first step. Recurrence is a property of conservative systems, which is a much narrower class than “systems in a bounded space”.

Why volume preservation is not an odd hypothesis #

The second condition sounds like a technical restriction and is in fact the defining property of the systems physics cares most about, which is why the theorem mattered rather than being a curiosity.

Newtonian mechanics, written in terms of positions and momenta, preserves volume in that space exactly — Liouville’s theorem, and it holds for any system with no friction and no external forcing. A region of possible states can be sheared, stretched and wound into a filament of enormous length, and its volume is unchanged at every instant. Nothing is lost and nothing is gained; the region only ever changes shape.

That is the precise sense in which such a system does not forget. A dissipative system collapses many different pasts onto the same present — which is why Newton’s basins can have three roots attracting whole regions — and a conservative one cannot, because collapsing volume to zero is exactly what it is forbidden to do.

The rotation at the top of this essay is the simplest possible example: it moves every arc to an arc of the same length. So is every reversible cellular automaton, and so, by construction, is the mechanics of a box of gas.

How long the wait is #

The return time is where the theorem stops being reassuring.

For the rotation above, the wait is manageable and computable. To come within $\varepsilon$ of the start takes about $1/\varepsilon$ steps, because the orbit is evenly spread at every stage and each new point covers its share. Halve the tolerance and roughly double the wait.

For anything with more dimensions the arithmetic is different in kind. If the state space has $d$ dimensions and the orbit visits it evenly, coming within $\varepsilon$ in every coordinate at once needs about $\varepsilon^{-d}$ steps. The exponent is the dimension, and dimensions in physical systems are counted in multiples of the particle count.

For a box of gas with $10^{23}$ molecules, the return time to a state resembling the starting one is of order $10^{10^{23}}$ — a number whose exponent has more digits than there are atoms in the observable universe. The theorem is true, the return is guaranteed, and no observation will ever see one.

That gap between “guaranteed” and “observable” is the whole content of the result for physics. A theorem with an unbounded waiting time makes a claim about eternity and no claim about any experiment.

The objection to the second law #

The theorem was used as an argument against statistical mechanics almost as soon as it appeared, and the resolution is worth following because it is a case of two correct statements looking incompatible.

Zermelo’s objection, 1896: gas molecules obey volume-preserving mechanics in a finite box, so by Poincaré’s theorem the gas must eventually return to any configuration it was in — including all the molecules bunched in one corner. Entropy would then have decreased. So the second law of thermodynamics cannot be exactly true.

The objection is valid. Boltzmann’s reply was not that it is wrong but that it is irrelevant: the recurrence time is the number above, the second law is a statement about what happens on any timescale anyone will observe, and a law that fails after $10^{10^{23}}$ seconds is not a law anyone needs to revise.

What makes this more than a debating point is the arithmetic. Both parties agreed about the mechanics and about the theorem; the disagreement was entirely about whether a statement with no time bound says anything. It does not, and recognising that is the same discipline as distinguishing a proof of existence from a method of construction.

Two ways of coming home #

This collection has one other guaranteed return, arrived at from the opposite direction, and setting them side by side isolates what each argument actually needs.

A random walk on a line or a plane comes home with probability one. That is a probabilistic theorem about a rule with no determinism in it, proved by summing a series, and it fails in three dimensions — a walker in space returns with probability about $0.34$ and otherwise wanders off forever.

Poincaré’s theorem is deterministic, says nothing about probability, and does not care about dimension at all. What it cares about is the two hypotheses: bounded space, preserved volume. The random walk on a line has neither — the line is unbounded — and returns anyway, for a reason that has nothing to do with volume.

So there are two independent routes to “it comes back”, and they agree on almost nothing else. The random walk’s return time has infinite expectation despite the probability being one; the recurrence time has a finite expectation given by Kac’s formula. The walk in three dimensions escapes; a conservative system in a bounded three-dimensional space cannot. The lesson is that “returns” is not one property but a conclusion several unrelated hypotheses happen to reach.

That figure is the case that shows volume preservation and reversibility are different conditions. The doubling map is measure-preserving and two-to-one, and the recurrence argument only used the first. Reversibility was needed for the pigeonhole version on a finite state space, not for this one.

Recurrence without a continuum #

The finite version of all this is not a lesser case; it is the version that applies to every computation.

The same bound, and the same uselessness. $2^{97}$ is about $10^{29}$, so the guarantee is that the pattern repeats within $10^{29}$ steps, and nothing observed in that figure will recur in any run anyone performs.

The waiting time also explains why the theorem is invisible in the figures rather than merely slow to appear. A return within $\varepsilon$ needs about $\varepsilon^{-d}$ steps, and the picture is a few dozen. Nothing on this page has been run long enough for recurrence to show, and nothing on this page could be. What the figures show is the hypotheses — a bounded space, a rule that moves things without squashing them — and the conclusion is a claim about a timescale no figure has.

That is an unusual position for this collection to be in, since the standing habit is that a figure should test what its caption claims. Here it cannot, and saying so is the honest version: these figures establish that the theorem’s conditions hold, and take the conclusion from the proof.

This is worth stating because it applies to every floating-point simulation on this site. A computer’s state is finite — sixty-four bits per number, finitely many numbers — so every simulated orbit is exactly periodic, including the chaotic ones. The period is astronomically long and depends on rounding rather than on the mathematics, so it is invisible; but a long chaotic run is, strictly, an eventually periodic sequence that has not yet closed. That is a second sense, alongside the shadowing problem, in which a computed orbit is not the orbit its caption names — and like the first, it is harmless for the question the figure is actually asked.

Returning is not the same as reversing #

There is a stronger-sounding statement nearby that the theorem does not support, and separating them is the difference between a real result and a paradox.

Recurrence says the state comes back near where it started. It does not say the system runs backwards to get there, or that the intervening history is undone in reverse order. The path from the starting region out and back is a forward path throughout; it simply happens to end up where it began.

This matters because the two get conflated in discussions of the arrow of time. A gas whose molecules re-bunch in a corner after $10^{10^{23}}$ seconds has not run backwards — it has run forwards through an unimaginable number of ordinary states and arrived, by an accident the theorem guarantees will eventually happen, at an unlikely one. The microscopic laws are reversible; the trajectory is not being reversed.

The distinction has a clean version in this field’s own terms. An irreversible rule like most cellular automata cannot satisfy the theorem’s hypothesis at all: it collapses states together, which is volume contraction in the finite setting, and the states it has collapsed are gone. A reversible rule keeps every state distinct forever, and recurrence follows immediately from counting.

So reversibility is the hypothesis and recurrence is the conclusion, and the conclusion does not restore the hypothesis’s direction. Knowing the system will return says nothing about the order in which it will pass through anything.

What the theorem is worth #

Given that the waiting time makes it unobservable, the value of the result is not predictive, and it is worth saying what it is instead.

It rules things out. Any claimed behaviour that a conservative system settles permanently into is wrong — no attractor, no final state, no permanent drift. That is a strong structural constraint and it is used constantly: it is why conservative and dissipative systems are studied with different tools, and why a strange attractor cannot exist in a Hamiltonian system.

It converts a physical question into a counting one. The interesting quantity is not whether a system returns but how long it takes, and the theorem makes that the only remaining question. The answer, per Kac’s formula, is that the mean return time to a region is inversely proportional to that region’s volume — a clean statement that turns “when” into a measurement. It also explains the $\varepsilon^{-d}$ above without any dynamics: a box of side $\varepsilon$ in $d$ dimensions has volume $\varepsilon^d$, and the return time is its reciprocal. The dimension enters through the geometry of small boxes, not through anything the rule does.

It is a pigeonhole argument in a place pigeonholes should not reach. The principle is usually taught as a fact about finite sets. Poincaré’s use of it establishes that the essential ingredient was never finiteness of count but finiteness of size — and once that is seen, the same argument works for how closely a fraction can approximate, for the rotation above, and for the gas in the box, which are otherwise three unrelated subjects.