Khinchin's Constant, or the Ergodic Heart of Continued Fractions

For almost every real number, the geometric mean of the coefficients in its continued fraction expansion converges to the same constant — roughly $2.6854520010$. This universality has nothing to do with any specific number and everything to do with a certain shift map on the unit interval being ergodic. That is the whole story; the rest is bookkeeping.

The elevator pitch

Take any real $x \in (0,1)$, say $\pi - 3 = 0.14159\ldots$. Compute $1/x$, take its integer part — that's the first coefficient $a_1$. Subtract, invert, take the integer part again — that's $a_2$. Iterate. You get a sequence $a_1, a_2, a_3, \ldots$ of positive integers that encode $x$ as

For $\pi - 3$ the sequence begins $7, 15, 1, 292, 1, 1, 1, 2, 1, 3, 1, 14, \ldots$. For $\sqrt{2}-1$ it is the boring $2, 2, 2, 2, \ldots$. For $e - 2$ it is the striking $1, 2, 1, 1, 4, 1, 1, 6, 1, 1, 8, \ldots$.

Now the question: is there anything universal we can say about these sequences? The arithmetic mean $(a_1 + \cdots + a_n)/n$ is a lost cause — for almost every $x$ it diverges, because occasional huge coefficients (a $292$ here, a $20776$ there) blow up the average. The geometric mean $(a_1 a_2 \cdots a_n)^{1/n}$ is more forgiving, because logs tame outliers. And here is the miracle: for almost every $x$, the geometric mean converges. To the same limit. Independent of $x$.

That limit is Khinchin's constant, denoted $K$. It is about $2.68545$, and it has a closed form as an infinite product we'll pin down at the end.

Refresher, then a shift of perspective

Any irrational $x \in (0,1)$ has a unique continued fraction expansion $[0; a_1, a_2, \ldots]$ with each $a_i \geq 1$ an integer. The classical results — best approximation, Hurwitz's theorem that infinitely many rationals $p/q$ satisfy $|x - p/q| < 1/(\sqrt 5 \, q^2)$, Lagrange's theorem that periodic expansions are exactly the quadratic irrationals — are beautiful, but they don't lead to Khinchin.

To reach Khinchin, we stop thinking of the expansion as an algorithm and start thinking of it as an orbit under a map. Define the Gauss map $T : [0,1) \to [0,1)$ by

In words: $T$ takes the fractional part of $1/x$. Notice what $T$ does to a continued fraction: if $x = [0; a_1, a_2, a_3, \ldots]$, then $1/x = a_1 + [0; a_2, a_3, \ldots]$, so $T(x) = [0; a_2, a_3, \ldots]$. The Gauss map is a shift on the coefficient sequence, and the first coefficient is exactly $a_1(x) = \lfloor 1/x \rfloor$. So $a_n(x) = a_1(T^{n-1} x)$. Computing the sequence of coefficients is the same as iterating $T$ from $x$ and reading off $\lfloor 1/y \rfloor$ each step. Every question about coefficient statistics becomes a question about orbit statistics, and orbit statistics are the province of ergodic theory.

The invariant measure

Iterating a map is only setup; the substance comes from a probability measure that $T$ preserves. Lebesgue measure does not work — $T$ smears the interval unevenly, piling mass near $0$ where the branches are steep. But Gauss himself, in a 1812 letter to Laplace, noticed that the measure

is invariant: $\mu(T^{-1} A) = \mu(A)$ for every Borel $A \subset [0,1)$. The density $\rho(x) = 1 / ((1+x) \log 2)$ is the Gauss-Kuzmin density. The $\log 2$ is a normalizer making $\mu([0,1)) = 1$; the shape $1/(1+x)$ is what carries the content.

The load-bearing check is direct. For $y \in [0,1)$, the pre-images of $y$ under $T$ are $\{1/(k+y) : k \geq 1\}$, one per positive integer. Change of variables from $y$ to $x = 1/(k+y)$ contributes $|dx/dy| = 1/(k+y)^2$. So $\rho$ is invariant if and only if

Plug in $\rho(x) = 1/((1+x)\log 2)$. The $k$-th term becomes $\frac{1}{\log 2} \cdot \frac{1}{(k+y)(k+1+y)}$. Partial fractions turn this into $\frac{1}{\log 2}\left(\frac{1}{k+y} - \frac{1}{k+1+y}\right)$, and the sum telescopes to $\frac{1}{\log 2} \cdot \frac{1}{1+y}$. That is exactly $\rho(y)$. The check is one line of algebra doing all the work.

The ergodic theorem does the rest

Given an invariant probability measure $\mu$, Birkhoff's ergodic theorem (1931) says: if $T$ is ergodic — meaning every $\mu$-invariant Borel set has $\mu$-measure $0$ or $1$ — then for every $\mu$-integrable $f$ and $\mu$-almost every $x$,

The time average along an orbit equals the space average against $\mu$. This is the entire engine.

Ergodicity of the Gauss map is a real theorem — it goes through a bounded-distortion estimate for cylinder sets $\{x : a_1(x) = k_1, \ldots, a_n(x) = k_n\}$ (the Renyi bound), or through showing $T$ is exact (Rokhlin, 1961). A careful account is in Einsiedler and Ward, Ergodic Theory with a View Towards Number Theory, Chapter 3. I am asserting ergodicity here rather than proving it; this is the one place the essay leans on an outside result.

Once ergodicity is granted, take $f(x) = \log \lfloor 1/x \rfloor$. Then $f(T^{k-1} x) = \log a_k(x)$, and Birkhoff says

In words: the log-average of the coefficients along a $\mu$-typical orbit equals the $\mu$-integral of $\log \lfloor 1/x \rfloor$. Exponentiating,

That right-hand side is Khinchin's constant. It exists because Birkhoff exists. It is universal because it depends on $\mu$ and $f$ but not on $x$.

Computing the constant

Break $[0,1)$ into the pieces where $\lfloor 1/x \rfloor = k$, i.e. $x \in [1/(k+1), 1/k)$. On each piece $\log \lfloor 1/x \rfloor = \log k$, so

Using $(k+1)^2 / (k(k+2)) = 1 + 1/(k(k+2))$ and dividing by $\log 2$ inside the log to turn it into $\log_2$, this rearranges to Khinchin's product:

The $k=1$ factor is trivial because $\log_2 1 = 0$; the series converges because $\log(1 + 1/(k(k+2))) = O(1/k^2)$. Numerically $K = 2.685452001065306\ldots$. Whether $K$ is even irrational is open. Nobody has ruled out $K = 5/2 + \varepsilon$ for a truly awkward $\varepsilon$.

What "almost every" really excludes

Almost every is measure-theoretic, and it excludes essentially every number you can write down with a formula. Quadratic irrationals are periodic (Lagrange), so their geometric means are the geometric means of the periodic block and only hit $K$ by numerical coincidence. $e$ has coefficients $[2;1,2,1,1,4,1,1,6,\ldots]$, dominated by ones, so its geometric mean tends to a value below $K$. Liouville numbers admit arbitrarily large jumps in $a_n$ and their geometric means blow up. Yet the union of all these exceptional sets has Lebesgue measure zero. Pick a real number uniformly from $(0,1)$ and it will obey Khinchin, with probability one. It's just that no number you can name is guaranteed to.

$\pi$ is conjectured — but not proved — to be one of the good numbers. Numerically the geometric mean of its first several billion partial quotients tracks $K$ closely, and that is the current state of the art. We do not even know that the partial quotients of $\pi$ are unbounded (it is widely believed but unproved). This is characteristic: the theorem is clean, the theorem's practical consequences for any specific irrational you care about are conjectures.

The moral

The classical view is that continued fractions are the best rational-approximation algorithm on the block. The ergodic view is that they are the orbit of $x$ under a shift, and their large-scale statistics are governed by the shift's invariant measure. Both views are true; only one produces Khinchin.

Whenever a numerical constant turns out to be independent of the input, the reason is almost always that the input has been quotiented out by a group action or a shift, and the constant is what remains — a functional of an invariant measure. Universality is what invariant measures look like from the outside. Khinchin's constant is the cleanest possible instance of the pattern: an infinite product, guaranteed by an ergodic theorem, that almost every real number quietly obeys.