29

Let $f(x)=\sum _{n=0}^{\infty } b_nx^n$ and $\frac{1}{f(x)}=\sum _{n=0}^{\infty } d_nx^n$. Then the coefficients of the reciprocal of $f(x)$ can be written down. The first few terms are:

$d_0 = \frac{1}{b_0}$,

$d_1 = -\frac{b_1}{b_0^2}$,

$d_2 = \frac{b_1^2-b_0 b_2}{b_0^3}$

$d_3 = -\frac{b_1^3-2 b_0 b_1 b_2+b_0^2 b_3}{b_0^4}$

...

I was wondering if there is a general recursive (preferably not of course) formula for the coefficients of the reciprocal. If an arbitrary $n$ is given, can I write down a formula for $d_n$ (recursive or not)?

Regards

//edit: as the comments below suggest I think people are misinterpretating the question. I am not looking for someone to show me how to solve a system of linear equations by substitution... I want a formula for d_n, Since posting the question, I found such a formula for $d_n$ at http://functions.wolfram.com/GeneralIdentities/7/, see the section on Ratios of the direct function ... if anyone knows of how this formula is derived or any other references to it or similar formulas please let me know... thanks

IV_
  • 1,063
AUK1939
  • 569
  • 1
    There's no $a_n$ in your functions, only $b_n$. And this really, really does look like homework, and is not research level.

    But anyway, before this gets closed: just use $f(x) \cdot \frac{1}{f(x)} = 1 + 0 \cdot x + 0 \cdot x^2 + \ldots$ and the formulae for multiplication of power series.

    – Zen Harper Jan 26 '11 at 18:59
  • See also: http://en.wikipedia.org/wiki/Power_series#Multiplication_and_division – Willie Wong Jan 26 '11 at 19:28
  • A homework question?? I dont think you read the question properly, or I lack your higher intellect. its one thing to solve a system on equations its quite another to write down a formula for the solution of the nth variable, n being arbitrary.

    I found the solution at http://functions.wolfram.com/GeneralIdentities/7/ ... I wonder how it was derived... Possibly somebody smart expanded enough terms found a pattern and used induction. That pattern is quite encryptic though. Any refs found would be appreciated.

    – AUK1939 Jan 26 '11 at 19:31
  • 1
    aukm, I agree: finding an explicit closed form formula for $d_n$ is not easy. However, by the method I outlined above, knowing the values of $d_0, d_1, \ldots, d_n$ (and the $b_j$), you can easily find $d_{n+1}$. Then you could easily programme a computer to calculate as many $d_j$ as you wish. As I understand it, that falls under the category of "recursive" formulae for $d_n$, which you say is allowed.

    But then you refer to a website listing explicit formulae. So that answers your question, doesn't it? If you want a proof or derivation instead, you should say this in your question!

    – Zen Harper Jan 26 '11 at 19:48
  • Zen, $d_n+1$ is found recursively by a computer but I am looking for a recursive or non recursive expression for $d_n+1$. A computer wont do this for me. I dont know how to make this any clearer. The formula in the link does answer my question, however, I would like to know how it was derived, who derived it ... basically where it came from. Note that it is a recursive formula. As far as I know a non recursive formula simply just doesn't exist. – AUK1939 Jan 26 '11 at 20:11
  • 1
    This is a straightforward application of Faa di Bruno (http://en.wikipedia.org/wiki/Fa%C3%A0_di_Bruno's_formula). Voting to close. – Qiaochu Yuan Jan 26 '11 at 21:26
  • 5
    Yuan, how can you close this question on the basis that its a staightforward application of the Faa di Bruno equation ... since 1) this equation is not known by many (read the third line of this paper, http://www.romanpress.com/MathArticles/FaaDiBruno.pdf). and 2) Even if the Faa di Bruno equation does provide an answer, its equivalence to the answer posted at http://functions.wolfram.com/GeneralIdentities/7/ is far from trivial ... and finally before you stupidly voted to close this question look at the responses below. The problem may not be hard but certainly not trivial!! – AUK1939 Jan 27 '11 at 00:11
  • 2
    And Willie Wong!! the reference you pointed me to simply suggests you didnt even read or understand the question ... so you foolishly closing the question has annoyed because I was having a good discussion about the solution with other members. – AUK1939 Jan 27 '11 at 00:12
  • 11
    Actually I don't see the need of closing this question, for these reasons: (1) even a simple question may be of interest to other professional mathematicians not in that very field; and (2) sometimes simple or naive questions here gave rise to wonderful answer by our best users. That said, I would recommend aukm not to feel offended, and to avoid quarreling --for some reasons it's considerd umpolite. – Pietro Majer Jan 27 '11 at 01:12
  • Related to http://mathoverflow.net/questions/238186/relating-face-polytopes-of-permutohedra-to-integer-partitions?lq=1 – Tom Copeland May 06 '16 at 22:21
  • See http://oeis.org/A263633, http://oeis.org/A133314, and http://oeis.org/A049019 – Tom Copeland Feb 01 '17 at 19:20
  • 13
    "This question is unlikely to help any future visitors"? NO, this question helped me a lot. – fairytale Jul 06 '17 at 16:58
  • This question has answers here: https://math.stackexchange.com/questions/1264615/inverse-rule-for-formal-power-series and here: https://math.stackexchange.com/questions/710252/multiplicative-inverse-of-a-power-series and here: https://mathoverflow.net/questions/53384/power-series-of-the-reciprocal-does-a-recursive-formula-exist-for-the-coeffic . – IV_ Dec 17 '19 at 18:55
  • 4
    Wow, this question helped me, how come it's too localized, very very weird!, If it should be closed, change the reason at least! – Rainb Jun 29 '21 at 05:56
  • The best answers are posted at https://math.stackexchange.com/a/4262414/945479 and https://math.stackexchange.com/a/4262417/945479. – qifeng618 Sep 28 '21 at 10:39

2 Answers2

24

Without loss of generality we can take $b_0$ to be 1, since \begin{equation*}\sum_{n=0}^\infty b_n x^n = b_0\biggl( 1+\sum_{n=1}^\infty (b_n/b_0)x^n\biggr). \end{equation*} Then for $b_0=1$ we have \begin{equation*} \frac1{f(x)} = \biggl( 1+\sum_{n=1}^\infty b_n x^n\biggr)^{-1}\\ =\sum_{m=0}^\infty (-1)^m\biggl( \sum_{n=1}^\infty b_n x^n\biggr)^m. \end{equation*} Expanding by the multinomial theorem and extracting the coefficient of $x^n$ gives \begin{equation*} \frac1{f(x)} = \sum_{n=0}^\infty \kern 3pt x^n \kern -5pt \sum_{m_1+2m_2+3m_3+\cdots = n} (-1)^{m_1+m_2+\cdots} \binom{m_1+m_2+\cdots}{m_1, m_2, \ldots} b_1^{m_1} b_2^{m_2}\cdots.\end{equation*}

Ira Gessel
  • 16,246
  • Cheers, again if you could provide a reference I would be grateful... – AUK1939 Jan 27 '11 at 01:06
  • 1
    If we generalize to $f(x)^r$ then the coefficients in the expansion are called potential polynomials (though they're expressed in terms of exponential generating functions, so the formula will have some additional factorials.) A formula for potential polynomials, which generalizes this formula, can be found in Comtet's Advanced Combinatorics, section 3.5. These polynomials are closely related to Bell polynomials. Another reference is Weiping Wang and Tianming Wang, General identities on Bell polynomials. Comput. Math. Appl. 58 (2009), no. 1, 104–118. – Ira Gessel Jan 28 '11 at 16:08
  • @IraGessel, I am curious about the Big-0 running time of potential polynomials . How can it be improved to nearly linear running time? Would Hansel Lifting help for this purpose? Thank you. – Frank Feb 18 '17 at 11:20
  • @Ira Gessel, I am curious about the Big-0 running time of potential polynomials . How can it be improved to nearly linear running time? Would Hansel Lifting help for this purpose? Thank you. – Frank Feb 18 '17 at 14:42
  • @Frank: I don't know. – Ira Gessel Feb 21 '17 at 02:16
  • The best answers are posted at https://math.stackexchange.com/a/4262414/945479 and https://math.stackexchange.com/a/4262417/945479. – qifeng618 Sep 28 '21 at 10:40
19

Assume $b_0=1$ to simplify things. You want a closed formula for the recursively defined sequence $$d_0=1$$ $$d_n=-\sum_{k=0}^{n-1}d_kb_{n-k}. $$ Let $\alpha=(\alpha_1,\dots,\alpha_r)\in \mathbb{N}_ +^\omega$ be a multi-index with length $l(\alpha):=r$ and weight $|\alpha|:=\sum_{j=1}^r\alpha_j$. Let's denote $b_\alpha:=b_{\alpha_1}\dots b_{\alpha_r}$.

We have (induction) $$d_n:=\sum_{|\alpha|=n}(-1)^{l(\alpha)}b_\alpha. $$

There are of course several equal terms in the sum, due to the commutativity; summing equal terms, a corresponding smaller set of indices would be the increasing multi-indices (the number of terms in the sum would then be the number of partitions $p(n)$).

Pietro Majer
  • 56,550
  • 4
  • 116
  • 260
  • Or you can derive it just expanding the formal series $(1+y)^{-1}$ with $y:=\sum_{k=1}^\infty b_k x^k$, and reordering. – Pietro Majer Jan 26 '11 at 21:17
  • 2
    Impressive, and ofcourse for $b_0 \neq 0$ the recursive relationship would be

    $d_n=\begin{cases} \frac{1}{b_0} & n=0 \ -\sum {k=0}^{n-1} \frac{d_k}{b_0}b{n-k} & n\geq 1 \end{cases}$

    Can you point me to any references. I am not familiar with multi indices, but will look into them to see how this recursive relation is solved. Also by any chance have you got any good references to solution techniques (like this one) to recurrence relations. This is a type of math I have never learnt, but it is catching my interest more and more.

    – AUK1939 Jan 26 '11 at 21:30
  • While far from being related to the actual problem, the discussion reminds me the following:

    (a) exact formula $$ det(A)=\sum_{\sigma \in S_n} \epsilon({\sigma}) {a_{1}}^{\sigma(1)} \cdots {a_{n}}^{\sigma(n)} $$ compared with

    (b) recursive formula: expansion by the first line of the determinant.

    Difficult to say if (a) or (b) can give a closed formula for say the determinant of a Hilbert matrix.

    – Luis H Gallardo Jan 26 '11 at 21:39
  • 3
    @aukm. Well, usually one treates recursive equations for sequences translating them into equations for the generating functions, because power series have a rich algebraic and analytic structure. For these techiques, you may like (some chapters of) Knuth's Concrete mathematics, or Wilf's Generating functionology. More advanced books, Enumerative Combinatorics by Stanley, and Analytic Combinatorics by Flajolet. – Pietro Majer Jan 26 '11 at 21:54
  • @Pietro Majer I am curious about the Big-0 running time of your generating functions which I encountered in a Rutgers mathematics for computer science course. How can it be improved to nearly linear running time? Would Hansel Lifting help for this purpose? Thank you – Frank Feb 18 '17 at 17:38
  • @Pietro Majer, Your generating formula for d subscript n has a closed form analytical solution rather than a recurrence relation making it parallelizable across n cores of a current many core CPU or GPU. Please confirm whether I am right or wrong. – Frank Feb 18 '17 at 23:36
  • Indeed one could possibly compute the $b_\alpha$ independently. I am not an expert of computer science however, so consider posting a new question. – Pietro Majer Feb 19 '17 at 09:14
  • @Pietro Majer, Thank you for your assertion that the b subscript n can be computed independently. Your comment is supported by the following URL which I just read, http://mathoverflow.net/questions/238186/relating-face-polytopes-of-permutohedra-to-integer-partitions?lq=1, – Frank Feb 19 '17 at 11:01
  • @Pietro Majer, Could we write a paper together on the idea of pararallelizing the polynomial division of Fourier spectra? Thank you. – Frank Feb 19 '17 at 11:04
  • 3
    (Incidentally, (7 years later) I see I forgot one author in a reference I gave above: Analytic Combinatorics by Flajolet and Sedgewick. My inexcusable excuses!) – Pietro Majer Jan 08 '19 at 17:02