Puede ser instructivo para demostrar este resultado a partir de primeros principios y resultados básicos, la explotación de las propiedades de cumulant funciones de generación (exactamente como en el estándar de pruebas del Teorema del Límite Central). Esto nos obliga a entender que la tasa de crecimiento de la generalizada armónica de los números $$H(n,s)=\sum_{k=1}^n k^{-s}$$ for $s=1, 2, \ldots.$ These growth rates are well-known and easily obtained by comparison to the integrals $\int_1^n x^{-s}dx$: they converge for $s \gt 1$ and otherwise diverge logarithmically for $s=1$.
Let $n \ge 2$ and $1 \le k \le n$. By definition, the cumulant generating function (cgf) of $(X_k - 1/k)/B_n$ is
$$\psi_{k,n}(t) = \log\mathbb{E}\left(\exp\left(\frac{X_k - 1/k}{B_n}t\right)\right) = -\frac{t}{k B_n} + \log\left(1 + \frac{-1 + \exp(t/B_n)}{k}\right).$$
The series expansion of the right hand side, obtained from the expansion of $\log(1+z)$ around $z=0$, takes the form
$$\psi_{k,n}(t) = \frac{(k-1)}{2 k^2 B_n^2}t^2 + \frac{k^2 - 3k + 2}{6 k^3 B_n^3} t^3 + \cdots + \frac{k^{j-1} - \cdots \pm (j-1)!}{j! k^j B_n^j}t^j + \cdots.$$
The numerators of the fractions are polynomials in $k$ with leading term $k^{j-1}$. Because the log expansion converges absolutely for $\left|\frac{-1 + \exp(t/B_n)}{k}\right| \lt 1$, this expansion converges absolutely when
$$\left|\exp(t/B_n) - 1\right| \lt k.$$
(In case $k=1$ it converges everywhere.) For fixed $k$ and increasing values of $n$, the (obvious) divergence of $B_n$ implies the domain of absolute convergence grows arbitrarily large. Thus, for any fixed $t$ and sufficiently large $n$, this expansion converges absolutely.
For sufficiently large $n$, then, we may therefore sum the individual $\psi_{k,n}$ over $k$ term by term in powers of $t$ to obtain the cgf of $S_n/B_n$,
$$\psi_n(t) = \sum_{k=1}^n \psi_{k,n}(t) = \frac{1}{2}t^2 + \cdots + \frac{1}{B_n^j}\left(\sum_{k=1}^n \left(k^{-1} - \cdots \pm (j-1)!k^{-j}\right)\right)\frac{t^j}{j} + \cdots.$$
Taking the terms in the sums over $k$ one at a time requires us to evaluate expressions proportional to
$$b(s,j) = \frac{1}{B_n^j}\sum_{k=1}^n k^{-s}$$
for $j \ge 3$ and $s=1, 2, \ldots, j$. Using the asymptotics of generalized harmonic numbers mentioned in the introduction, it follows easily from
$$B_n^2 = H(n,1) - H(n,2) \sim \log(n)$$
that
$$b(1,j) \sim (\log(n))^{1-j/2}\to 0$$
and (for $s \gt 1$)
$$b(s,j) \sim (\log(n))^{-j/2}\to 0$$
as $n$ grows large. Consequently all terms in the expansion of $\psi_n(t)$ beyond $t^2$ converge to zero, whence $\psi_n(t)$ converges to $t^2/2$ for any value of $t$. Since convergence of the cgf implies convergence of the characteristic function, we conclude from the Levy Continuity Theorem that $S_n/B_n$ approaches a random variable whose cgf is $t^2/2$: that is the standard Normal variable, QED.
This analysis uncovers just how delicate the convergence is: whereas in many versions of the Central Limit Theorem the coefficient of $t^j$ is $O(n^{1-j/2})$ (for $j \ge 3$), here the coefficient is only $O(((\log(n))^{1-j/2})$: the convergence is much slower. In this sense the sequence of standardized variables "just barely" becomes Normal.
We can see this slow convergence in a series of simulations. The histograms display $10^5$ independent iterations for four values of $n$. The red curves are graphs of standard normal density functions for visual reference. Although there is evidently a gradual tendency towards normality, even at $n=1000$ (where $(\log(n))^{-1/2} \approx 0.38$ is still sizable) there remains appreciable non-normality, as evidenced in the skewness (equal to $0.35$ in this sample). (It is no surprise the skewness of this histogram is close to $(\log(n))^{-1/2}$, because that's precisely what the $t^3$ plazo en la cgf).
Aquí es el R
código para aquellos a quienes les gusta experimentar.
set.seed(17)
par(mfrow=c(1,4))
n.iter <- 1e5
for(n in c(30, 100, 300, 1000)) {
B.n <- sqrt(sum(rev((((1:n)-1) / (1:n)^2))))
x <- matrix(rbinom(n*n.iter, 1, 1/(1:n)), nrow=n, byrow=FALSE)
z <- colSums(x - 1/(1:n)) / B.n
hist(z, main=paste("n =", n), freq=FALSE, ylim=c(0, 1/2))
curve(dnorm(x), add=TRUE, col="Red", lwd=2)
}