3
$\begingroup$

I have functions $A, B, F, S$ that are zero on $(-\infty, 0)$. And I have successfully represented the below equation as convolution and multiplication:

$\int_0^t {dt_1} \int_0^t {dt_2} B(t - t_2)F(t_2 - t_1)S(t - t_1)F(t_1)$

$=\int_0^t {dt_1}F(t_1) S(t - t_1)\int_{t_1}^t {dt_2} B(t - t_2)F(t_2 - t_1)$

$=\int_0^t {dt_1}F(t_1) S(t - t_1)\int_0^{t-t_1} {dt'_2} B(t - t_1 - t'_2)F(t'_2)$

$=\int_0^t {dt'_1}F(t - t'_1) S(t'_1)\int_0^{t'_1} {dt'_2} B(t'_1 - t'_2)F(t'_2)$

$=\int_0^t {dt'_1}F(t - t'_1) S(t'_1)(B*F)(t'_1)$

$=(F * (S(B*F)))(t)$

where $(F*G)(t)=\int_0^tF(t-t')G(t')dt'$. The first equality holds because $F(t_2-t_1)=0$ for $t_2<t_1$. The second equality results by substitution $t'_2=t_2-t_1$. The third equality results by substitution $t'_1=t-t_1$.

It doesn't look like it can be represented as a convolution at first, but actually it turns out that it is represented by a combination of convolution and multiplication of functions.

But I failed to represent the below equation as a combination of convolution and multiplication. Is it possible, or not? If it is not possible, how can one prove that it is not possible?

$\int_0^t {dt_1} \int_0^t {dt_2} \int_0^t {dt_3} B(t - t_3)F(t_3 - t_2)S(t - t_2)F(t_2 - t_1)F(t_1)$

For me, it doesn't look like it is possible. But like the equation in the first line which doesn't seem to be represented as convolution at first glance, it may be possible. If anyone succeeds, please let me know. Thank you.


Is it possible to represent the following function as a convolution?

$$ \int_0^tdt_1 \int_0^tdt_2 \int_0^tdt_3 A(t-t_3)F(t_3-t_2)S(t-t_2)F(t_2-t_1)S(t_3-t_1)F(t_1) $$

$\endgroup$
4
  • 2
    $\begingroup$ Do not write "functions $B(x), F(x), S(x)$ that are zero for $x<0$" -- write "functions $B,F,S$ that are zero on $(-\infty,0)$" instead. Do not write $B(t)*F(t)$ --write $(B*F)(t)$ instead. $\endgroup$ Commented Mar 10, 2021 at 14:44
  • 1
    $\begingroup$ @IosifPinelis, while I certainly agree with you about what usage is proper, I think that the usage in the parent post is so familiar that, while technically improper, it is not really any more objectionable than writing $\frac{\mathrm d}{\mathrm dx}(x^2)$ instead of something like the more technically correct but horrible $\bigl(D(x \mapsto x^2)\bigr)(x)$. Certainly I have seen the notation in the post used by mathematicians wiser and abler than I. $\endgroup$
    – LSpice
    Commented Mar 10, 2021 at 14:48
  • 2
    $\begingroup$ @LSpice : (i) I have certainly seen expressions such as "function $f(x)$", especially in older papers. (ii) In your example, the "correct" expression is indeed horrible, whereas the expression $\frac d{dx}(x^2)$ is not really objectionable at all, since it is easy and natural to attach a meaning to the derivative of an expression (rather than of a function). (iii) In my suggestions, in contrast with that example, the correct versions are, not only not horrible, but they are also neater, more compact, and easier to understand than the incorrect ones, especially $B(t)*F(t)$ vs. $(B*F)(t)$. $\endgroup$ Commented Mar 10, 2021 at 15:05
  • $\begingroup$ @IosifPinelis I have changed the notation :) $\endgroup$
    – J.G. Kang
    Commented Mar 11, 2021 at 11:43

1 Answer 1

3
$\begingroup$

$\newcommand\R{\mathbb R}$ Most likely, the integral $$I:=\int_0^t dt_1 \int_0^t dt_2 \int_0^t dt_3\, B(t-t_3)F(t_3-t_2) S(t-t_2)F(t_2-t_1)F(t_1)$$ cannot be expressed in terms of products and convolutions -- because the finite interval $[0,t]$ is not a group (or even a semigroup), and the restrictions $t_2\in[0,t]$ and $t_3\in[0,t]$ are unjustified, since the integrand contains no factors depending only on $t_2$ or only on $t_3$. Since the functions $B,F,S$ are zero on $(-\infty,0)$, the correct(ed) form of the integral $I$ seems to be \begin{align*} \tilde I&:=\int_\R dt_1 \int_\R dt_2 \int_\R dt_3\, B(t-t_3)F(t_3-t_2) S(t-t_2)F(t_2-t_1)F(t_1) \\ &=\int_\R {dt_2}\, (B*F)(t-t_2)S(t-t_2)(F*F)(t_2) \\ &=\Big(\big((B*F)S\big)*(F*F)\Big)(t), \end{align*} so that $\tilde I$ is expressed in terms of products and convolutions.


The similarly corrected integral added later to the OP, \begin{align*} J&:=\int_\R dt_1 \int_\R dt_2 \int_\R dt_3\, A(t-t_3)F(t_3-t_2)S(t-t_2)F(t_2-t_1)S(t_3-t_1)F(t_1), \end{align*} probably cannot be expressed in terms of products and convolutions.

A way to check it may be by the following shift observation, which addresses a comment by LSpice:

Suppose we want to show that a function $f$ cannot be expressed in terms of products and convolutions of functions $f_1,\dots,f_n$. To do this, we build all possible binary trees as follows. The nodes of such a tree are nonempty subsets of the set $[n]=\{1,\dots,n\}$. To each node $J$ there corresponds an expression in terms of products and/or convolutions of the functions $(f_j\colon j\in J)$. The root of the tree is the set $[n]$. Each node is split into two branches, whose roots are nonempty subsets, which become the nodes next up the tree. Each node is marked either by $p$ (for "product") or $c$ (for "convolution"), depending on whether the daughter expressions of the node get multiplied or convolved.

Then the functions $f_1,\dots,f_n$ get shifted horizontally, starting with the leaves of the tree. If two adjacent leaves are the branches of a $p$-node, their corresponding functions get shifted by the same real number, say $u$. After that, the two leaves get deleted and their parent node is marked by the same shift $u$. If two adjacent leaves are the branches of a $c$-node, their corresponding functions get shifted by two arbitrary real numbers, say $u_1$ and $u_2$. After that, the two leaves get deleted and their parent node is marked by shift $u_1+u_2$. This process then continues down to the root. For every $p$-node encountered during this process, the shifts assigned to the daughter branches get equalized, say by imposing the corresponding equality constraint. Finally, at the root of the tree, we check if the function $f$ has the resulting shift property.

As a comparatively simple case, suppose that a function $f$ can be expressed in terms of products and convolutions of functions $f_1,\dots,f_n$ as follows: \begin{equation} f=\Big(\prod_{j\in J_1}f_j\Big)*\cdots*\Big(\prod_{j\in J_m}f_j\Big), \tag{1} \end{equation} where $\{J_1,\dots,J_m\}$ is a partition of the set $\{1,\dots,n\}$. Suppose next that, for each $k=1,\dots,m$ and each $j\in J_k$, the function $f_j$ is shifted horizontally by a real number $u_k$. Then $f$ gets shifted by $u_1+\dots+u_m$.

For any given functions $f_1,\dots,f_n$, there may be only finitely many representations of $f$ in the form (1); so, we can try and exhaust all of them.


By considering all the distinguishable trees, it is not hard to see that the integral $J$ cannot be expressed in terms of products and convolutions. This can be done even with zero shifts, and even with the additional assumption that $A=S=F$, as shown in this Mathematica notebook; you can also see the notebook's pdf image.

A few comments on that notebook:

  1. E.g., the possible representation $c(F,p(c(F,F),c(F,p(F,F))))$ of the integral $J$ (with $A=S=F$) is the result of the application of the sequence $(c, p, c, c, p)$ of operations $c$[=convolution] or $p$[=product] to the (unlabeled) tree $(x, ((x, x), (x, (x, x))))$, where $x$ represents any leaf. In our case (concerning the integral $J$), there are $6$ trees (each with $6$ leaves) and $2^5$ sequences of operations $c$ or $p$. So, there are $6\times2^5=192$ possible representations of the integral $J$ in terms of products and convolutions.

  2. For positive real $K$ and $s$, let $f_{K,s}:=Kf_s$, where $f_s$ is the density function of the normal distribution with mean $0$ and variance $s^2$. Then for any positive real $K_1,K_2,s_1,s_2$ we have $$f_{K_1,s_1}*f_{K_2,s_2}=f_{K,s}\quad\text{for}\quad K=K_1K_2,\ s=\sqrt{s_1^2+s_2^2},$$ $$f_{K_1,s_1}f_{K_2,s_2}=f_{K,s}\quad\text{for}\quad K=\frac{K_1K_2}{\sqrt{2\pi}\sqrt{s_1^2+s_2^2}},\ s=\frac{s_1s_2}{\sqrt{s_1^2+s_2^2}}.$$ So, the family $(f_{K,s})$ is closed with respect to operations $c$ and $p$.

  3. So, taking e.g. $F=f_{1,1}$, we see that each possible representation of the integral $J$ in terms of products and convolutions (such as e.g. $c(F,p(c(F,F),c(F,p(F,F))))$) is of the form $f_{K,s}$ and thus is determined by the corresponding pair $(K,s)$ of positive real numbers. The integral $J$ for $F=f_{1,1}$ is also of the form $f_{K,s}$, actually with $(K,s)=(\frac{1}{8 \pi ^{3/2}},\sqrt{2})$.

  4. The direct calculation done in the mentioned Mathematica notebook shows that none of the $192$ pairs $(K,s)$ corresponding to the $192$ possible representations of the integral $J$ in terms of products and convolutions equals $(\frac{1}{8 \pi ^{3/2}},\sqrt{2})$. So, the integral $J$ cannot be expressed in terms of products and convolutions.

$\endgroup$
17
  • 2
    $\begingroup$ It seems like there must be some sort of mechanical procedure for formally converting expressions of the form in the original post to expressions involving only convolution and multiplication, when this is possible. Do you know any easy criterion that identifies when it is not possible? $\endgroup$
    – LSpice
    Commented Mar 10, 2021 at 17:07
  • 2
    $\begingroup$ @LSpice : I think this is a good comment, but I don't know of such an algorithm or a criterion of whether it would succeed. $\endgroup$ Commented Mar 10, 2021 at 18:32
  • $\begingroup$ Thanks for your valuable answer. Also, I came up with another integral representation which I added in parent post. It doesn't seem possible. Like @LSpice mentioned, I'd love to have criteria or something that one can tell it is possible to represent as convolution or not $\endgroup$
    – J.G. Kang
    Commented Mar 11, 2021 at 11:44
  • 1
    $\begingroup$ @LSpice : Now your comment has been addressed: There is a way to identify at least some (maybe all) situations when such integrals cannot be expressed in terms of products and convolutions. $\endgroup$ Commented Mar 11, 2021 at 16:53
  • 2
    $\begingroup$ I'm not sure your shift observation quite works for this problem. For instance with the form $S(F*G)$ one needs to translate $F,G$ by $u$ and $S$ by $2u$ in order to translate the entire form by $2u$. The problem seems to be more subtle than this - it should not be possible to express this form via convolution, but this is because a certain graph underlying the form is "crossing" rather than "non-crossing" in nature which is preventing it from splitting into lower multilinearity forms. Unfortunately have to run, may explain later. $\endgroup$
    – Terry Tao
    Commented Mar 11, 2021 at 17:05

You must log in to answer this question.

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