4

Let $f$ and $g$ be sequences of functions and $F$ and $G$ their corresponding generating functions, \begin{eqnarray*} & F(z)=\sum_{n\in\mathbb{N}} f_n z^n \\ & G(z)=\sum_{n\in\mathbb{N}} g_n z^n \\ \end{eqnarray*} The termwise multiplication of these sequences is an operation called Hadamar product, $$F \cdot G (z) := \sum_{n\in\mathbb{N}} f_ng_n z^n $$ As it can be found in Wikipedia, Hadamard products and diagonal generating functions, it states that $$ \boxed{F \cdot G (z) = \frac{1}{2\pi} \int_{0}^{2\pi} F\big(\sqrt{z} e^{it}\big) G\big(\sqrt{z} e^{-it}\big) dt}.$$

How do you get this formula? I have thoroughly looked for it without finding any derivation, in Wikipedia article there is a cite to a book, but this formula does not seem to appear in any form there.


This question is related to:

Algorithm for computing Hadamard product of two rational generating functions,

Does having closed forms of two generating functions guarantee that one can find the closed form of their term-by-term product? and

Is there an easier formulation for the Hadamard product of certain pair of series? as well.

Dr Potato
  • 772

2 Answers2

2

By Cauchy product, $$F(\sqrt{z} x)\, G(\sqrt{z}/x) = \sum_{m=0}^\infty f_m (\sqrt{z})^m \sum_{k=0}^\infty g_k (\sqrt{z})^k \, x^{m-k} $$ Use the Cauchy integral representation of the Kronecker delta (1 if m=k, 0 otherwise) $$\frac{1}{2\pi\,i}\oint \frac{dx}{x} x^{m-k} = \delta_{m,k} $$ Thus $$\frac{1}{2\pi\,i}\oint \frac{dx}{x} F(\sqrt{z} x)\, G(\sqrt{z}/x) = \sum_{m=0}^\infty f_m \, g_m z^m $$ Now make the substitution $x=e^{i\,t}.$ Thus we get the OP's answer, corrected by the leading factor of $1/(2 \pi)$,

$$ \frac{1}{2 \pi} \int_{0}^{2 \pi} F(\sqrt{z} e^{it})\, G(\sqrt{z} e^{-i\,t}) dt = \sum_{m=0}^\infty f_m \, g_m z^m $$

user321120
  • 6,740
  • In this answer, the radius $R$ of the contour is not specified. I have a question: Can $R$ be arbitrarily big/small? Even does it returns the same value after taking its limit to $\infty$ or $0$, correspondingly? – Dr Potato Jun 17 '20 at 03:13
  • @DrPotato I assumed that $F$ and $G$ were sufficiently well-behaved. I believe as long as R is less that the nearest singularity of $F$ or $G,$ this should work. – user321120 Jun 18 '20 at 01:51
  • What if one or both of the generating functions have an accumulation of singularities at zero? Should I calculate the residue at $0$ only? – Dr Potato Jun 19 '20 at 04:09
  • @DrPotato Sounds like something like $e^{1/z}?$ Ask the entire MSE your exact application as a new question, and also ping me on it. I usually look at MSE every other day or so. – user321120 Jun 19 '20 at 22:23
  • Actually it is asked in Value of the contour integral around an interesting singularity of $1/(e^z-1) \cdot 1/(e^{1/z}-1)$.. As you can see, it results from a Hadamard termwise product and although it is already answered, does not say anything about taking the limit as the radius vanishes. I conjecture we can not take the limit but we can evaluate the contour integral for some positive radius, and I WONDER IF THIS STILL IS A VALID FORMULA FOR THE DIRECT PRODUCT of the two explicit generating functions. – Dr Potato Jun 24 '20 at 18:28
1

Two derivations from operstional calculus:

For

$$A(x) = \sum_{n \geq 0} a_n x^n$$

and

$$\widetilde{A}(x) = \sum_{n \geq 0} a_n \frac{x^n}{n!} = e^{a.x}$$

with $(a.)^n = a_n$, the Hadamard product is given by

$$\sum_{n \geq 0} a_n x^n \frac{D_{x=0}^n}{n!} G(x)= \sum_{n\geq 0} a_ng_n x^n $$

with $d/dx= D_x$,

or more concisely,

$$\widetilde{A}(:xD_{x=0}:)G(x)= \exp(a.:xD_{x=0}:)G(x)=G(a.x)= (A*G)(x)$$

with $:xD_x:^n = x^nD_x^n$, by definition, a notational convenience.

The derivatives may be coded as the Cauchy contour integrals

$$g_n = \frac{D^n_{z=0}}{n!}G(z) = \frac{1}{2\pi i} \oint_{|z|<\epsilon} \frac{G(z)}{z^{n+1}}dz$$

where $\epsilon$ is less than the radii of the circles of convergence of the two series.

So, with appropriate changes of variables,

$$H(x)= (F*G)(x)$$

$$ = \frac{1}{2\pi i} \sum_{n \geq 0} f_n x^n \oint_{|z|<\epsilon} \frac{G(z)}{z^{n+1}}dz.$$

$$= \frac{1}{2\pi i} \oint_{|z|<\epsilon} \frac{F(\frac{x}{z})G(z)}{z}dz$$

$$= \frac{1}{2\pi i} \oint_{|z|<\alpha} \frac{F(\frac{\sqrt{x}}{z})G(z\sqrt{x})}{z}dz$$

$$= \frac{1}{2\pi} \int_0^{2\pi} F(\sqrt{x}\alpha^{-1}e^{-it})G(\sqrt{x}\alpha e^{it})dt$$

$$= \frac{1}{2\pi} \int_{0}^{2\pi} F(\sqrt{x}e^{-it})G(\sqrt{x}e^{it})dt,$$

assuming both series reps are convergent for $\alpha=1$. The last real integral is convergent for all functions bounded in the segment of integration.

For some discussion of the validity of these formulas, see "Hadamard grade of power series" by Allouche and France.

Alternatively, note (cf. this MSE answer)

$$\exp(txD_x)f(x)=f(e^t x).$$

Then

$$\frac{1}{2\pi} \int_{-\pi}^{\pi} F(ue^{-it})G(ve^{it})dt$$

$$= \frac{1}{2\pi} \int_{-\pi}^{\pi} e^{-ituD_u}e^{itvD_v} dt F(u)G(v)$$

$$= \frac{1}{2\pi} \int_{-\pi}^{\pi} e^{-it(uD_u-vD_v)} dt F(u)G(v)$$

$$=\frac{sin[\pi(uD_u-vD_v)]}{\pi(uD_u-vD_v)}F(u)G(v)$$

$$= \sum_{j,k \geq 0} sinc(\pi(j-k)) f_j g_k u^jv^k$$

$$= \sum_{k \geq 0} f_k g_k (uv)^k.$$

Tom Copeland
  • 2,416
  • I will read it carefully. I have another question on how to calculate the direct product using convolution. Please take a look at it here – Dr Potato Apr 22 '20 at 03:59
  • 1
    Consider https://en.m.wikipedia.org/wiki/Parseval%27s_theorem. – Tom Copeland Apr 22 '20 at 04:15
  • There is another more general formula as a Mellin convolution that I've derived and presented in a draft at my miniarxiv. – Tom Copeland May 13 '20 at 05:29
  • Please share link. – Dr Potato May 15 '20 at 16:02
  • I think the Hadamard product for this example does not exist. – Dr Potato Jun 10 '20 at 22:59
  • I got confused by ...what is $A$ doing at the begining? and then totally lost by the " : " notation, at around the third part of what you wrote. Really tried to understand, but not used to such sort of umbral calculus or whatever notation ...got dizzy. Could you please explain? – Dr Potato Jun 14 '20 at 20:49
  • @DrPotato, wade through this: https://math.stackexchange.com/questions/191752/whats-umbral-calculus-about/1085165#1085165 The umbral notation and colons simplify expressions. You don't have to use $\sum$, binomial coefficients, or divided powers $x^n/n!$ and indices all the friggin time. To check your understanding replace the umbral notation with the cumbersome standard notation and $:AB:^n$ with $A^nB^n$. – Tom Copeland Jun 14 '20 at 22:09
  • I have a question: if the radius of convergence of both series is infinite, can $\epsilon$ be arbitrarily big/small? Is the value the same after taking its limit to $\infty$ or $0$, correspondingly? – Dr Potato Jun 17 '20 at 03:11
  • @DrPotato, I believe so. Study complex analysis and the insights it offers on analytic functions and the classic Cauchy Integral. Very useful and interesting. – Tom Copeland Jun 17 '20 at 04:46
  • I also believe so, but... suppose both generating functions are holomorphic in the complex plane, and that the sum of (even the absolute value of) all their residues is finite. How to show that this Hadamard product formula is independent from the radius $R$? – Dr Potato Jun 17 '20 at 17:09