The Basel problem (first posed in 1644) is the challenge of summing reciprocal squares,

\[\sum_{k=1}^\infty\frac{1}{k^2} = \frac{1}{1^2} + \frac{1}{2^2} + \frac{1}{3^2} + \cdots = \,\,?\]

The first person to solve it was Euler. His strategy is over-represented in expository math blogs, but I can’t resist writing about it and describing a small extension Euler made.

The Basel problem

Euler started by representing the $\sin$ function as an infinite product, something like a factorisation of the Taylor series. We can write a polynomial $p(x)$ with roots ${a_1, a_2, \ldots, a_n}$ as

\[p(x) = C(x - a_1)(x - a_2)\cdots(x - a_n) = C\prod_{i = 1}^n(x - a_i)\]

for some constant $C$. Similarly, we can try to write $\sin(x)$ as a product of terms corresponding to its zeros at $\pi k \in \mathbb{Z}$:

\[\begin{align*} \sin(x) & = C \prod_{k\in\mathbb{Z}}(x - \pi k) \\ & = C x \prod_{k\geq 1}(x - \pi k)(x + \pi k) \\ & = C x \prod_{k\geq 1}(x^2 - \pi^2 k^2). \end{align*}\]

At this point, we run into a problem: this expression will never converge for any finite $x$, since the factors blow up as $|k| \to \infty$. To stop this from happening, we divide the $k$th (quadratic) term through by $-\pi^2k^2$. This gives us factors which approach $1$ as $k \to \infty$ (the infinite product version of the divergence test) without changing the roots. (We imagine lumping all of these terms into $C$.) Now, divide through by $x$ and then send $x \to 0$:

\[\lim_{x\to 0}\frac{\sin(x)}{x} = 1.\]

This tells us we should set $C = 1$ (after dividing out the problematic terms). So, we have the heuristic relation

\[\frac{\sin(x)}{x} = \prod_{k\geq 1}\left(1 - \frac{x^2}{\pi^2 k^2}\right).\]

It turns out this identity is true, and can be proved using complex analysis. In general, convergence of these infinite products is a delicate matter, and we ignore it completely in what follows.

Euler’s ingenious trick was to identify the coefficient of $x^2$ on either side. On the left, the Taylor series for $\sin(x)/x$

\[\frac{\sin(x)}{x} = \sum_{n\geq 0} \frac{(-1)^n x^{2n}}{(2n+1)!} = 1 - \frac{x^2}{6} + \mathcal{O}(x^4)\]

gives a coefficient $-1/6$. On the right, we can formally extend the result for polynomials

\[\prod_{i=1}^n (1 - a_i x) = 1 - x\sum_{i=1}^n a_i + \mathcal{O}(x^2)\]

to infinite series to obtain the coefficient

\[-\sum_{k\geq 1} \frac{1}{\pi^2k^2}= -\frac{1}{\pi^2}\sum_{k\geq 1} \frac{1}{k^2}.\]

Equating gives Euler’s answer to the Basel problem,

\[\sum_{k=1}^\infty \frac{1}{k^2} = \frac{\pi^2}{6}.\]

Cool! We can obtain the same result in other ways, notably the cotangent method, the Fourier series for $f(x) = x$ and an elementary “Proofs from the Book” method, but Euler’s approach is the original and (I think) the best.

Euler’s trick for higher terms

At this point, it’s natural to try the same trick for higher terms in the Taylor series. Let’s look at $x^4$. On the left, we have

\[\frac{(-1)^2}{5!} = \frac{1}{120}.\]

On the right, we have something more complicated. Let’s clean up our notation, and define

\[\prod_{i\geq 1}(1 + a_i x) = \sum_{n\geq 0}A_n x^n.\]

We have already used the result

\[A_1 = \sum_{i\geq0}a_i.\]

Generally, $A_n$ will be a sum of products of $n$ distinct $a_i$, since each will contribute a factor of $x$ and occurs by choosing the $x$ term in the corresponding factors and the $1$ term in all remaining factors. Conversely, this is the only way to get an $x^n$ term. Hence,

\[A_n = \sum_{i_1 < \cdots < i_n} a_{i_1}\cdots a_{i_n}.\]

This is complete and general, but hard to work with. Let’s repackage $A_2$ in a useful way. We have

\[A_2 = \sum_{i < j} a_ia_j = \frac{1}{2}\sum_{i \neq j} a_ia_j = \frac{1}{2}\left(\sum_{i, j} a_ia_j - \sum_{i}a_i^2\right) = \frac{1}{2}\big(A_1^2 - A_1^{(2)}\big),\]

where $A_1^{(k)} = \sum_i a_i^k$. Finally, we go back to evaluating the coefficient of $x^4$ on the right hand side. Since $a_k = -1/(\pi k)^2$ and $A_1 = -1/6$, the coefficient on the right is

\[\begin{align*} \frac{1}{2}\big(A_1^2 - A_1^{(2)}\big) & = \frac{1}{2}\left(\frac{1}{36} - \frac{1}{\pi^4}\sum_{k\geq 0}\frac{1}{k^4}\right). \end{align*}\]

Setting this equal to $1/120$ and rearranging, we obtain

\[\sum_{k\geq 0}\frac{1}{k^4} = \frac{\pi^4}{90}.\]

We could go on in a similar way, producing ad hoc recurrences for $A_n$ in terms of smaller $A_m$ and $A_1^{(n)}$. In fact, Euler did so, calculating coefficients all the way up to $x^{26}$, but this ad hoc method is tedious and inefficient.

It would be nice if one could find an expression for $A_n$ (which in our case is known) in terms of $A_1^{(k)}$ (which we would like to know). We would be hoping for something like the recurrence satisfied by the Bernoulli numbers, since with the benefit of hindsight, we know that $A_1^{(k)} \propto B_{2k}$. Sadly, the author has yet to undertake this endeavour!