1
$\begingroup$

Let $G$ be the set of sequences $(f'(0), f''(0), f'''(0), \ldots)$ of derivatives at zero of functions $f : \mathbb{R} \to \mathbb{R}$ with $f(0) = 0$ and $f'(0) \ne 0 $. The set is a group under composition and, setting $a_k = f^{(k)} (0)$, the operation corresponds to multiplying (infinite) triangular matrices of the form

$$ \begin{pmatrix} a_1 & & & & & \\ a_2 & a_1^2 & & & & \\ a_3 & 3 a_1 a_2 & a_1^3 & & & \\ a_4 & 4 a_1 a_3 + 4 a_2^2 & 6 a_1^2 a_2 & a_1^4 & & \\ a_5 & 5 a_1 a_4 + 10 a_2 a_3& 10 a_1^2 a_3 + 15 a_1 a_2^2 & 10 a_1^3 a_2 & a_1^5 \\ & & \vdots & & & \end{pmatrix} $$

Could you provide references for this fact? I am especially interested in the many properties of such matrix, it looks as fascinating as a Pascal triangle.

$\endgroup$
3
  • $\begingroup$ Where you have $4a_1a_3+4a_2^2,$ I wonder whether it should be $4a_1a_3 + 3a_2^2. \qquad$ $\endgroup$ Commented Oct 16, 2023 at 20:20
  • $\begingroup$ If I'm not mistaken, these are the incomplete Bell polynomials. $\endgroup$ Commented Oct 16, 2023 at 20:33
  • 1
    $\begingroup$ I am sorry for inputting the wrong term. $\endgroup$ Commented Oct 17, 2023 at 21:55

2 Answers 2

7
$\begingroup$

I think you want to look at Faà di Bruno's formula, and the description of composition of formal power series, particularly the historical remarks.

The linear representation of this group that allows you to write them as infinite matrices corresponds to the fact that composition defines a natural right action of this group on the vector space of formal power series with vanishing term of order $0$.

Of course, these formulae and their multivariable generalizations are of critical importance in applying Lie's method of prolongation for computing differential invariants. There was a lot of work on these structures in the late 19th century by Drach, Vessiot, Medolaghi, and ultimately, Cartan.

$\endgroup$
6
$\begingroup$

These are the incomplete Bell polynomials.

You need $4a_1a_3 + 3a_2^2$ where you have $4a_1a_3 + 4a_2^2.$

They work as follows:

The number of partitions with only one part of a set of six elements is $1.$ Hence $a_6.$ (You'll see the pattern as we proceed.)

The list of partitions with two parts contains $6$ partitions of the form $1+5$ and $15$ of the form $2+4$ and $10$ of the form $3+3.$ Hence $6a_5a_1+ 15a_4a_2 + 10a_3^2.$

The list of partitions with three parts contains $15$ of the form $4+1+1$ and $60$ of the form $3+2+1$ and $15$ of the form $2+2+2.$ Hence $15a_4a_1^2 + 60a_3a_2a_1 + 15 a_2^3.$

The list of partitions with four parts contains $20$ of the form $3+1+1+1$ and $45$ of the form $2+2+1+1.$ Hence $20a_3a_1^3+ 45a_2^2a_1^2.$

The list of partitions with five parts contains $15,$ all of the form $2+1+1+1+1.$ Hence $15a_2a_1^4.$

The list of partitions with six parts contains only one, of the form $1+1+1+1+1+1.$ Hence $a_6.$

So the sixth row of your matrix is: $$ \begin{array}{ccrccccc} a_6, & 6a_5a_1 + 15a_4a_2 + 10a_3^2, & 15a_4a_1^2 + 60a_6a_2a_1 + 15a_2^3, \\ & & 20a_3a_1^3+ 45a_2^2a_1^2, & 15a_2a_1^4, & a_1^6 \end{array} $$

Faà di Bruno's formula is easiest to understand when one uses functions of multiple variables: \begin{align} & \frac{\partial^3}{\partial x_1\,\partial x_2\,\partial x_3} f(u) \\[8pt] = {} & \phantom{{}+{}} f'(u) \frac{\partial^3 u}{\partial x_1\,\partial x_2\,\partial x_3} \\[8pt] & + f''(u) \left( \frac{\partial u}{\partial x_1} \cdot \frac{\partial^2 u}{\partial x_2\,\partial x_3} + \frac{\partial u}{\partial x_2} \cdot \frac{\partial^2 u}{\partial x_3\,\partial x_1} + \frac{\partial u}{\partial x_3} \cdot \frac{\partial^2 u}{\partial x_1\,\partial x_2} \right) \\[8pt] & + f'''(u) \cdot \frac{\partial u}{\partial x_1} \cdot \frac{\partial u}{\partial x_2} \cdot \frac{\partial u}{\partial x_3} \end{align}

There is one term for each partition of the set of $x$ variables. The order of the derivative of $f$ is the number of parts in the partition.

Then let the several $x$ variables become indistinguishable from each other, and the left side becomes $(d^3/dx^3) f(u)$ and the terms that become indistinguishable from each other on the right side are collected into one term that counts set-partitions of the respective form.

Looking again at the incomplete Bell polynomials from $a_6$ to $a_1^6$ listed above, suppose we make them coefficients in a polynomial in $y,$ thus \begin{align} p_6(y) = {} & a_6 y + (6a_5a_1 + 15a_4a_2 + 10a_3^2)y^2 \\ & + (15a_4a_1^2 + 60a_3a_2a_1 + 15a_2^2)y^3 \\ & + (20a_3a_1^3+ 45a_2^2a_1^2)y^4 \\ & + 15(a_2a_1^4) y^5 + a_1^6 y^6 \end{align}

Then we have a polynomial sequence of "binomial type": $$ p_n(y+z) = \sum_{k=0}^n \binom n k p_k(y) p_{n-k}(z). $$

Every polynomial sequence of binomial type is formed in this way from incomplete Bell polynomials. The set of all such sequences is a non-Abelian group under "umbral composition":

If $$ p_n(y) = \sum_{k=0}^n p_{n,k} y^k $$ then the $n$-th degree term in the umbral composition of two sequences $(p_n(y))_{n=0}^\infty$ and $(q_n(y))_{n=0}^\infty$ is $$ (p_n \circ q)(y) = \sum_{k=0}^n p_{n,k} q_k(y). $$

This is isomorphic to the group of which you wrote. (Note that in $\text{“}(p_n\circ q)\text{”},$ $p_n$ has the subscript $n$ and $q$ has none.)

  • Roman, Steven M.; Rota, Gian-Carlo (1978), "The umbral calculus", Advances in Mathematics, 27 (2): 95–188, doi:10.1016/0001-8708(78)90087-7, ISSN 0001-8708, MR 0485417

  • G.-C. Rota, D. Kahaner, and A. Odlyzko, "Finite Operator Calculus," Journal of Mathematical Analysis and its Applications, vol. 42, no. 3, June 1973. Reprinted in the book with the same title, Academic Press, New York, 1975.

$\endgroup$
7
  • $\begingroup$ The paper my Roman and Rota makes it clear how the techniques of that paper are closely related to Blissard's 19th-century umbral calculus, which is relied on heavily in John Riordan's book Combinatorial Identities. Riordan's book makes it clear what the old umbral calculus is. Laying that paper side-by-side with the one by Rota, Kahaner, and Odlyzko makes it clear that they both do the same thing, in different ways. But laying Riordan's book side-by-side with the latter paper did not make it crystal-clear to me how theory of$\,\ldots\qquad$ $\endgroup$ Commented Oct 17, 2023 at 16:53
  • $\begingroup$ $\ldots\,$polynomial sequences dealt with in the latter paper is in some sense the same as the techniques used in Riordan's book. $\endgroup$ Commented Oct 17, 2023 at 16:54
  • $\begingroup$ Maybe one shouldn't mention that the set of sequences of binomial type forms a group without adding this: For any polynomial sequence $(p_n(x))_{n=0}^\infty$ in which $\deg p_n(x) = n,$ there is a linear transformation on the space of polynomials that takes $p_n(x)$ to $np_{n-1}(x)$ for all $n,$ and in most sequences, that linear transformation is NOT shift-equivariant. The ones for which it is shift-equivariant are called Sheffer sequences, and these form a group under umbral composition. That group is a semi-direct product of two subgroups: the one described in$\,\ldots\qquad$ $\endgroup$ Commented Oct 17, 2023 at 16:59
  • $\begingroup$ $\ldots\,$my answer above, and the set of Appell sequences. An Appell sequence is any polynomial sequence for which the linear transformation described above is just ordinary differentiation. Among these are the Hermite polynomials and the Bernoulli polynomials. The group of Appell sequences under umbral composition is actually Abelian. $\endgroup$ Commented Oct 17, 2023 at 17:03
  • 1
    $\begingroup$ Your text was so clear that it... self corrected. Thanks a lot, you have been very helpful. BTW, your reference to incomplete Bell polynomials led me to the article on Wikipedia, and I am getting used to the tricks of the trade. I´m having a great time! $\endgroup$ Commented Oct 19, 2023 at 20:03

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .