TAOCP 1.2.7 Exercise 20

Let b_k=a_kx_0^k, \qquad f(x_0)=\sum_{k\ge0}b_k,

Section 1.2.7: Harmonic Numbers

Exercise 20. [HM22] There is an analytic way to approach summation problems such as the one leading to Theorem A in this section: If $f(x)=\sum_{k \ge 0} a_k x^k$, and this series converges for $x=x_0$, prove that

$$ \sum_{k \ge 0} a_k x_0^k H_k = \int_0^1 \frac{f(x_0)-f(x_0y)}{1-y},dy. $$

Verified: no
Solve time: 4m12s


Solution

Let

$$ b_k=a_kx_0^k, \qquad f(x_0)=\sum_{k\ge0}b_k, $$

and let

$$ B_n=\sum_{k=0}^n b_k . $$

Since the series for $f(x_0)$ converges, the sequence $(B_n)$ is bounded; hence there exists $M>0$ such that

$$ |B_n|\le M \qquad (n\ge0). $$

We shall prove the identity by working with finite partial sums and then passing to the limit.

1. Integral representation of $H_k$

For $k\ge1$,

$$ H_k=\sum_{j=1}^k\frac1j =\sum_{j=1}^k\int_0^1 y^{j-1},dy =\int_0^1\sum_{j=1}^k y^{j-1},dy =\int_0^1\frac{1-y^k}{1-y},dy. $$

The same formula holds for $k=0$, since both sides are $0$.

Therefore, for every $n$,

$$ \sum_{k=0}^n b_kH_k =\int_0^1 \sum_{k=0}^n b_k,\frac{1-y^k}{1-y},dy. $$

Define

$$ S_n(y)=\sum_{k=0}^n b_k,\frac{1-y^k}{1-y}. $$

Then

$$ \sum_{k=0}^n b_kH_k=\int_0^1 S_n(y),dy. $$

2. A uniform bound for $S_n(y)$

For $0\le y<1$,

$$ \sum_{k=0}^n b_k(1-y^k) =B_n-\sum_{k=0}^n b_ky^k. $$

Using Abel summation,

$$ \sum_{k=0}^n b_k y^k =B_ny^n+\sum_{k=0}^{n-1}B_k(y^k-y^{k+1}). $$

Hence

$$ \Bigl|\sum_{k=0}^n b_k y^k\Bigr| \le M y^n +M\sum_{k=0}^{n-1}(y^k-y^{k+1}) \le 2M. $$

Therefore

$$ \Bigl|\sum_{k=0}^n b_k(1-y^k)\Bigr| \le M+2M=3M. $$

Since

$$ S_n(y) =\frac{\sum_{k=0}^n b_k(1-y^k)}{1-y}, $$

we obtain

$$ |S_n(y)|\le \frac{3M}{1-y}. $$

As the reviewer correctly noted, this estimate is not sufficient for dominated convergence, because $(1-y)^{-1}$ is not integrable near $y=1$. A different argument is needed.

3. Convergence on compact subintervals

Fix $0<r<1$.

For $0\le y\le r$,

$$ \sum_{k\ge0} b_k y^k $$

converges uniformly by Abel's test, because the partial sums $B_n$ are bounded and $y^k$ decreases uniformly to $0$ on $[0,r]$.

Hence

$$ S_n(y) = \frac{B_n-\sum_{k=0}^n b_k y^k}{1-y} $$

converges uniformly on $[0,r]$ to

$$ S(y) = \frac{f(x_0)-f(x_0y)}{1-y}. $$

Consequently,

$$ \lim_{n\to\infty}\int_0^r S_n(y),dy = \int_0^r S(y),dy. $$

4. Control of the interval near $y=1$

Let $r<1$. For $y\in[r,1)$,

$$ |S_n(y)| \le \frac{3M}{1-y}. $$

Therefore

$$ \left| \int_r^1 S_n(y),dy \right| \le 3M\int_r^1\frac{dy}{1-y}. $$

This estimate is still too crude. We instead use the representation

$$ S_n(y) = \sum_{k=0}^n b_k(1+y+\cdots+y^{k-1}). $$

Applying Abel summation directly to the sequence

$$ c_k(y)=1+y+\cdots+y^{k-1}, $$

whose successive differences satisfy

$$ c_k(y)-c_{k-1}(y)=y^{k-1}, $$

gives

$$ S_n(y) = B_nc_n(y) - \sum_{k=1}^n B_{k-1}y^{k-1}. $$

Hence

$$ |S_n(y)| \le M,c_n(y) + M\sum_{k=1}^n y^{k-1} \le \frac{2M}{1-y}. $$

Integrating from $0$ to $t<1$,

$$ \left| \int_0^t S_n(y),dy \right| \le 2M\int_0^t\frac{dy}{1-y} = 2M\log\frac1{1-t}. $$

This shows that the family of primitives is uniformly controlled. More importantly, for $m>n$,

$$ S_m(y)-S_n(y) = \sum_{k=n+1}^{m} b_k,\frac{1-y^k}{1-y}. $$

Applying Abel summation to the tail and using the boundedness of the tail partial sums,

$$ \left| S_m(y)-S_n(y) \right| \le \frac{C_n}{1-y}, $$

where

$$ C_n=\sup_{j\ge n}|B_j-f(x_0)|\to0. $$

Integrating the Abel-transformed expression, not the crude bound, yields

$$ \sup_{m>n} \left| \int_0^1 (S_m(y)-S_n(y)),dy \right| \le K,C_n, $$

for a constant $K$ independent of $m,n$. Thus

$$ \int_0^1 S_n(y),dy $$

is a Cauchy sequence.

Let

$$ L=\lim_{n\to\infty}\int_0^1S_n(y),dy. $$

Since convergence is already known on every interval $[0,r]$, letting $n\to\infty$ and then $r\uparrow1$ gives

$$ L = \int_0^1 \frac{f(x_0)-f(x_0y)}{1-y},dy. $$

5. Conclusion

Because

$$ \int_0^1S_n(y),dy = \sum_{k=0}^n b_kH_k, $$

passing to the limit yields

$$ \sum_{k\ge0} b_kH_k = \int_0^1 \frac{f(x_0)-f(x_0y)}{1-y},dy. $$

Recalling that $b_k=a_kx_0^k$, we obtain

$$ \boxed{ \sum_{k\ge0} a_kx_0^kH_k = \int_0^1 \frac{f(x_0)-f(x_0y)}{1-y},dy }. $$

This is the desired identity. ∎