Let $v=\sum_{n=1}^\infty(-1)^n/n^4$ ($v$ for "value"), let $S=(\sum_{n=1}^m(-1)^n/n^4)_{m\in\mathbb Z_{\ge1}}$ be the partial sums, and let $e=(|S_n-v|)_{n\in\mathbb Z_{\ge1}}$ be the errors. Also note $v=-\eta(4)=-7\pi/720$ (Dirichlet eta).$^1$ My computations$^2$ of the first few errors suggest that the errors are strictly decreasing: $e_1>e_2>e_3>\dots.$ How can we prove this?

(I came up with this conjecture while solving a homework exercise for my undergraduate numerical analysis class. My conjecture is NOT part of the homework; I am asking out of curiosity. Also note $v=\operatorname{Li}_4(-1)$ (polylogarithm), if that helps.)

$^1$The proof that $\eta(4)=7\pi^4/720,$ or that $v=-7\pi^4/720,$ can be reduced to the proof that $\zeta(4)=\pi^4/90.$ For instance, let $u=\sum_{n=1}^\infty n^{-4}.$ Then $$v+u=\sum_{n=1}^\infty(-1)^nn^{-4}+\sum_{n=1}^\infty n^{-4}=\sum_{n=1}^\infty((-1)^n+1)n^{-4}=2\sum_{n=1}^\infty(2n)^{-4}=u/8.$$ Since $u=\pi^4/90$, we thus have $v=-7u/8=-7\pi^4/720.$ (Proof taken (almost) verbatim from ProofWiki.)

$^2$I used Matlab to compute the following errors. For example, $\operatorname{round}_4(e_1)=5.30\!\cdot\!10^{-2},\operatorname{round}_5(e_2)=9.53\!\cdot\!10^{-3},\operatorname{round}_5(e_3)=2.81\!\cdot\!10^{-3},$ where for any real number $x$ and any integer $n\ge0,$ $\operatorname{round}_n(x)$ is $x$ rounded to $n$ decimals.

First 150 errors: $5.30\!\cdot\!10^{-2},9.53\!\cdot\!10^{-3},2.81\!\cdot\!10^{-3},1.09\!\cdot\!10^{-3},5.07\!\cdot\!10^{-4},2.65\!\cdot\!10^{-4},1.51\!\cdot\!10^{-4},9.27\!\cdot\!10^{-5},5.98\!\cdot\!10^{-5},4.02\!\cdot\!10^{-5},2.81\!\cdot\!10^{-5},2.02\!\cdot\!10^{-5},1.49\!\cdot\!10^{-5},1.12\!\cdot\!10^{-5},8.57\!\cdot\!10^{-6},6.68\!\cdot\!10^{-6},5.29\!\cdot\!10^{-6},4.24\!\cdot\!10^{-6},3.44\!\cdot\!10^{-6},2.81\!\cdot\!10^{-6},2.33\!\cdot\!10^{-6},1.94\!\cdot\!10^{-6},1.63\!\cdot\!10^{-6},1.38\!\cdot\!10^{-6},1.18\!\cdot\!10^{-6},1.01\!\cdot\!10^{-6},8.71\!\cdot\!10^{-7},7.56\!\cdot\!10^{-7},6.58\!\cdot\!10^{-7},5.76\!\cdot\!10^{-7},5.07\!\cdot\!10^{-7},4.47\!\cdot\!10^{-7},3.96\!\cdot\!10^{-7},3.52\!\cdot\!10^{-7},3.14\!\cdot\!10^{-7},2.81\!\cdot\!10^{-7},2.52\!\cdot\!10^{-7},2.27\!\cdot\!10^{-7},2.05\!\cdot\!10^{-7},1.86\!\cdot\!10^{-7},1.68\!\cdot\!10^{-7},1.53\!\cdot\!10^{-7},1.39\!\cdot\!10^{-7},1.27\!\cdot\!10^{-7},1.17\!\cdot\!10^{-7},1.07\!\cdot\!10^{-7},9.81\!\cdot\!10^{-8},9.03\!\cdot\!10^{-8},8.32\!\cdot\!10^{-8},7.68\!\cdot\!10^{-8},7.10\!\cdot\!10^{-8},6.58\!\cdot\!10^{-8},6.10\!\cdot\!10^{-8},5.66\!\cdot\!10^{-8},5.27\!\cdot\!10^{-8},4.90\!\cdot\!10^{-8},4.57\!\cdot\!10^{-8},4.27\!\cdot\!10^{-8},3.99\!\cdot\!10^{-8},3.73\!\cdot\!10^{-8},3.49\!\cdot\!10^{-8},3.27\!\cdot\!10^{-8},3.07\!\cdot\!10^{-8},2.89\!\cdot\!10^{-8},2.71\!\cdot\!10^{-8},2.56\!\cdot\!10^{-8},2.41\!\cdot\!10^{-8},2.27\!\cdot\!10^{-8},2.14\!\cdot\!10^{-8},2.02\!\cdot\!10^{-8},1.91\!\cdot\!10^{-8},1.81\!\cdot\!10^{-8},1.71\!\cdot\!10^{-8},1.62\!\cdot\!10^{-8},1.54\!\cdot\!10^{-8},1.46\!\cdot\!10^{-8},1.39\!\cdot\!10^{-8},1.32\!\cdot\!10^{-8},1.25\!\cdot\!10^{-8},1.19\!\cdot\!10^{-8},1.13\!\cdot\!10^{-8},1.08\!\cdot\!10^{-8},1.03\!\cdot\!10^{-8},9.80\!\cdot\!10^{-9},9.35\!\cdot\!10^{-9},8.93\!\cdot\!10^{-9},8.53\!\cdot\!10^{-9},8.15\!\cdot\!10^{-9},7.79\!\cdot\!10^{-9},7.45\!\cdot\!10^{-9},7.13\!\cdot\!10^{-9},6.83\!\cdot\!10^{-9},6.54\!\cdot\!10^{-9},6.27\!\cdot\!10^{-9},6.01\!\cdot\!10^{-9},5.76\!\cdot\!10^{-9},5.53\!\cdot\!10^{-9},5.31\!\cdot\!10^{-9},5.10\!\cdot\!10^{-9},4.90\!\cdot\!10^{-9},4.71\!\cdot\!10^{-9},4.53\!\cdot\!10^{-9},4.36\!\cdot\!10^{-9},4.19\!\cdot\!10^{-9},4.04\!\cdot\!10^{-9},3.89\!\cdot\!10^{-9},3.74\!\cdot\!10^{-9},3.61\!\cdot\!10^{-9},3.48\!\cdot\!10^{-9},3.35\!\cdot\!10^{-9},3.23\!\cdot\!10^{-9},3.12\!\cdot\!10^{-9},3.01\!\cdot\!10^{-9},2.91\!\cdot\!10^{-9},2.81\!\cdot\!10^{-9},2.71\!\cdot\!10^{-9},2.62\!\cdot\!10^{-9},2.54\!\cdot\!10^{-9},2.45\!\cdot\!10^{-9},2.37\!\cdot\!10^{-9},2.29\!\cdot\!10^{-9},2.22\!\cdot\!10^{-9},2.15\!\cdot\!10^{-9},2.08\!\cdot\!10^{-9},2.02\!\cdot\!10^{-9},1.95\!\cdot\!10^{-9},1.89\!\cdot\!10^{-9},1.83\!\cdot\!10^{-9},1.78\!\cdot\!10^{-9},1.72\!\cdot\!10^{-9},1.67\!\cdot\!10^{-9},1.62\!\cdot\!10^{-9},1.57\!\cdot\!10^{-9},1.53\!\cdot\!10^{-9},1.48\!\cdot\!10^{-9},1.44\!\cdot\!10^{-9},1.40\!\cdot\!10^{-9},1.36\!\cdot\!10^{-9},1.32\!\cdot\!10^{-9},1.28\!\cdot\!10^{-9},1.25\!\cdot\!10^{-9},1.21\!\cdot\!10^{-9},1.18\!\cdot\!10^{-9},1.15\!\cdot\!10^{-9},1.12\!\cdot\!10^{-9},1.09\!\cdot\!10^{-9},1.06\!\cdot\!10^{-9},1.03\!\cdot\!10^{-9},1.00\!\cdot\!10^{-9},9.74\!\cdot\!10^{-10}$

  • $\begingroup$ The is a well know result about alternating series, such as the one you are studying, that says that $|s-\sum^n_{k=1}(-1)^{k+1}a_k|<a_{n+1}$. Here $a_n\searrow0$. One can also fins the exact error and see is deacreasing. (See Apostol's analysis book, second edition pp 188-189.) $\endgroup$ Sep 20 '20 at 17:03

In terms of the Hurwitz zeta function your $m$th error is \begin{align*} \left| {\sum\limits_{n = m + 1}^\infty {\frac{{( - 1)^n }}{{n^4 }}} } \right| & = \left| { \sum\limits_{n = 0}^\infty {\frac{{( - 1)^n }}{{(n + m + 1)^4 }}} } \right| \\ &= \frac{1}{{16}}\left( {\zeta \left( {4,\tfrac{{m + 1}}{2}} \right) - \zeta \left( {4,\tfrac{m}{2} + 1} \right)} \right) \\ & = \frac{1}{{96}}\int_0^{ + \infty } {t^3 e^{ - (m + 1)t/2} \frac{{1 - e^{ - t/2} }}{{1 - e^{ - t} }}dt} , \end{align*} showing that the error is strictly decreasing with respect to $m$.


Partial answer

$v$ is an alternating series. It is a general result regarding alternating series, that the error of the partial sum is bounded by the next general term. $\frac{1}{(n+1)^4}$ in this specific case.

See wikipedia article for more details and a proof of this result.

Also for an alternating series $$S= \sum_{n=0}^\infty \frac{(-1)^n}{a_n}$$ the partial sums $$S_k= \sum_{n=0}^k \frac{(-1)^n}{a_n}$$ are such that $\{S_{2k}\}$ is an increasing sequence. Hence the error $\{e_{2k}\}$ is a decreasing sequence. Similarly $\{S_{2k+1}\}$ is a decreasing sequence and $\{e_{2k+1}\}$ is also a decreasing sequence.

In our case we have

$$\begin{cases} e_{2k+1} &= S_{2k+1} - S = \sum_{n=2k+2}^\infty \frac{(-1)^n}{a_n}\\ e_{2k} &= S -S_{2k} = -\sum_{n=2k+1}^\infty \frac{(-1)^n}{a_n} \end{cases}$$

Therefore $$\begin{aligned}e_{2k}-e_{2k+1} &= 2S - S_{2k} - S_{2k+1}\\ &= \frac{1}{(2k+1)^4} - 2 \sum_{n=2k+2}^\infty \frac{(-1)^n}{n^4}\\ &= \frac{1}{(2k+1)^4} - 2 \sum_{n=k+1}^\infty \left(\frac{1}{(2n)^4} - \frac{1}{(2n+1)^4}\right) \end{aligned}$$

Remains to prove that this sequence is non negative.

Which is done here by Martin R.

  • $\begingroup$ Also that the errors are decreasing. $\endgroup$ Sep 20 '20 at 15:50
  • $\begingroup$ @mathcounterexamples.net How does this answer my question? Do we have $n^{-4}\le e_n$ or something? $\endgroup$ Sep 20 '20 at 15:54
  • $\begingroup$ You have $e_n < 1/(n+1)^4$. $\endgroup$ Sep 20 '20 at 15:54
  • $\begingroup$ How do you get from $e_n<1/(n+1)^4$ to $\dots<e_{n+2}<e_{n+1}<e_n<\dots<e_1$? $\endgroup$ Sep 20 '20 at 15:56
  • 2
    $\begingroup$ @mathcounterexamples.net Be careful, working with alternating series abstractly will not suffice here (on its own). For instance, you might ask: Are the errors (strictly) decreasing in the series $1-1+2^{-1}-2^{-1}+3^{-1}-3^{-1}+\dots$? The answer is "no": The even indices $n$ have the property $e_n=0.$ $\endgroup$ Sep 20 '20 at 16:13

Using the approach from Proving that $\frac{1}{(2k+1)^4} - 2 \sum_{n=k+1}^\infty \left(\frac{1}{(2n)^4} - \frac{1}{(2n+1)^4}\right)$ is non negrative we have the following general result:

If $(a_n)$ is a decreasing and convex sequence with $\lim_{n\to \infty} a_n = 0$ then the absolute values of series remainders (error terms) of the alternating series $\sum_n (-1)^n a_n$ are decreasing .

If the sequence is strictly convex then the absolute values of the series remainders are strictly decreasing.

A sequence is convex if $a_{n+1} \le (a_{n} + a_{n+2})/2$ for all $n$. This is satisfied in our case because $f(x) = 1/x^4$ is a convex function.

Proof of the above statement: Let $$ e_n = \left | \sum_{j=n}^\infty (-1)^j a_j \right|= a_n - a_{n+1} + a_{n+2} - a_{n+3} + a_{n+4} - \ldots $$ denote the absolute value of the $n^\text{th}$ series remainder. Then $$ e_{n+1} - e_{n} = -a_n +2 a_{n+1} -2 a_{n+2} +2 a_{n+3} - 2 a_{n+4} + \ldots\\ = (-a_n +2 a_{n+1} - a_{n+2}) + (- a_{n+2} +2 a_{n+3} - a_{n+4}) + \ldots \le 0 $$ from the convexity condition.


This is from Apostol's book on Analysis:

Suppose $a_n\searrow0$ and $s=\sum^\infty_{n=1}a_n$, and $s_n=\sum^n_{k=1}(-1)^{k+1}a_k$. Then $$\begin{align} a_{n+1}-a_{n+2}&<\sum^\infty_{k=1}(a_{n+2k-1}-a_{n+2k})\\ &=(-1)^n(s-s_n)=a_{n+1}-\sum^\infty_{k=1}(a_{n+2k}-a_{n+2k+1})<a_{n+1} \end{align} $$

Maybe this can be used to show the decay in errors that you are observing for $a_n=\frac{1}{n^4}$. Certainly $|s-s_n|<a_{n+1}$ so at each step, the bounds improve.

  • $\begingroup$ See xFioraMstr18's comment above: The error terms are not necessarily decreasing for an alternating series with decreasing absolute values. $\endgroup$
    – Martin R
    Sep 20 '20 at 17:44
  • $\begingroup$ @MartinR: do you have an example with strictly decreasing values ? $\endgroup$
    – user65203
    Sep 20 '20 at 18:16
  • $\begingroup$ @YvesDaoust: Unfortunately not. I only cited a comment which I found relevant here, but I did not search myself for pathological examples. $\endgroup$
    – Martin R
    Sep 20 '20 at 19:19

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.