I would like the function $f(n,M)$, where $n$ and $M$ are integers and $n\le M$, so that $f$ satisfies the following two conditions:

(1) $\sum_{n=0}^M (-1)^n f(n,M) = \tfrac{1}{2}$

(2) $f(n,M) \approx 1$ whenever $n\ll M$

The function is needed as a summation mollifier to ensure a piecewise function is continuous (each region of the piecewise function is a series expansion). The function $f(n,M) = e^{-\frac{k n^2}{M^2}}$ is close for large $k$, but the summation is not exactly $\frac{1}{2}$ and thus the piecewise function has a discontinuity.

I should also mention that $f(n,M) = 1$ for $n<M$ and $f(n,M) = 1/2$ for $n=M$ works, but I would prefer a smoother function.

Below is some additional context to address Gottfried Helms's comment.

I am computing a finite series approximation for $\frac{1}{1+g(x)}$ for smooth $g$ where $0<g(x)<1$ for $x<1$ and $1<g(x)$ for $x>1$.

For $x<1$, I would like to expand $\frac{1}{1+g(x)} = \sum_{n=0}^\infty (-1)^n g(x)^n$.

For $x>1$, I would like to expand $\frac{1}{1+g(x)} = \frac{g(x)^{-1}} {1+g(x)^{-1}} = g(x)^{-1} \sum_{n=0}^\infty (-1)^n g(x)^{-n}$.

Terms for both series converge to the alternating sequence $\sum_{n=0}^\infty (-1)^n$ as $x\rightarrow 1$. I want to introduce a mollifier $f(n,M)$ so that the $M$-term approximation to the function is continuous:

$\frac{1}{1+g(x)} \approx \begin{cases} \hphantom{g(x)^{-1}}\sum_{n=0}^M (-1)^n f(n,M) g(x)^n & x<1 \\ g(x)^{-1} \sum_{n=0}^M (-1)^n f(n,M) g(x)^{-n} & x \ge 1 \end{cases}$. The two requirements for $f(n,M)$ are intended to ensure that (1) the approximation is continuous at $x=1$ where $\frac{1}{1+g(1)} = 1/2$ and (2) the approximation is accurate far from the point $x=1$ where $g(x)$ is not near 1.

I am not very knowledgeable about divergent series or series acceleration, so I may be missing something easier. The reason I am leaning toward modifying the terms is that I would like to preserve the polynomial structure of the approximation in $g(x)$ or $g(x)^{-1}$.

  • $\begingroup$ Hmm, if I understand things correctly it should be of vital interest how much the growth rate of the coefficients is which should be mollified (arithmetic/geometric/hypergeometric growth). The matrix given by Gerry Myerson can mollify geometric series with q=-2 at most. I think what you ask for is the transformation-matrix for Eulersummation (optimally: with configurable order). Is that correct so far? Second question: do you want to mollify the sequence of partial sums or the coefficients themselves? $\endgroup$ Oct 1, 2014 at 18:35
  • $\begingroup$ I should mention that I have a satisfactory solution, so please don't spend too much time on this. Thank you for the information though. $\endgroup$ Oct 1, 2014 at 19:43
  • $\begingroup$ I used a van Wijngaarden transformation and it worked extremely well. $\endgroup$ Oct 1, 2014 at 20:35
  • $\begingroup$ Very well; I use Euler-summation with configurable order. The order must be adapted: when the series to mollify is $1-1+1-1...$ order 1 is needed, when it is $1-3+9-27+81-...+...$ order 3 is needed and so on. If I see it correctly the current proposal is the matrix (or a matrix very near to that) for Euler-summation of order 1. But after your last formula it seems that you can circumvent quotients of absolute values $ \gt 1$, so that matrix should sufffice for all cases and no additional consideration is needed. $\endgroup$ Oct 1, 2014 at 23:57

2 Answers 2


I'm not sure what exactly "smoother" means in the context of a function defined on a discrete set, but maybe this will do you: $f(0,M)=1-2^{-(M+1)}$, $f(n,M)=1-3\times2^{-(M-n+2)}$ for $n\ge1$. In tabular form, $$\matrix{1/2&&&&\cr3/4&1/4&&&\cr7/8&5/8&1/4&&\cr15/16&13/16&5/8&1/4&\cr31/32&29/32&13/16&5/8&1/4\cr}$$ where the rows are for $M=0,1,\dots$ and the columns for $n=0,1,\dots$.

  • $\begingroup$ By "smoother", I mean that $|f(n,M)-f(n+1,M)|$ is small. I like your function; it avoids the annoying optimization step in my function. $\endgroup$ Oct 1, 2014 at 0:21

In case anyone stumbles on this question, the function $f(n,M) = \cos(\pi x n / 2M)$ works well, where $x$ is close to 1. In the limit $M \rightarrow \infty$, using $x=1$ is exact. For finite $M$, $x$ is slightly less than 1 (found by optimization).


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

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