The set of $\mathcal{L}^p$ functions is closed under addition, multiplication and convolution operations. I'd like to know an example of a set of functions which is closed under convolution and not under addition and multiplication.
A family of functions which is closed under convolution operation but not under addition and multiplication. operations
-
2The easiest example is the standard orthonormal basis in $\ell^2(\mathbb{N})$, I presume. – 2011-07-27
-
0@Theo Buehler : I do not know of such a thing, could you please give a link or explain. – 2011-07-27
-
1@Rajesh: See http://en.wikipedia.org/wiki/Sequence_space. – 2011-07-27
-
0Let $e_n$ be the sequence that is zero except at place $n$. The convolution $c=(c_n) = a \ast b$ of two sequences $a=(a_n)$ and $b=(b_n)$ is given by $c_n = \sum_{k} a_k \cdot b_{n-k}$. Using this formula you have $e_{m+n} = e_m \ast e_n$. But obviously the set $\{e_m\} \subset \ell^p$ is neither closed under addition nor multiplication. This convolution arises when looking at the Fourier transform $\mathcal{F}(f\cdot g) = \mathcal{F}(f) \ast \mathcal{F}(g)$ of periodic functions, for example. [See e.g. here](http://en.wikipedia.org/wiki/Convolution#Discrete_convolution). – 2011-07-27
-
0@Theo Buehler : thank you for the example. I wonder there are any examples of set of functions of the form $f: \mathbb{R} \to \mathbb{R}$. – 2011-07-27
-
0I knew of it as a discrete time signals in signal processing. – 2011-07-27
-
0Exactly, that's one major application. I guessed so from your notation, that's why I only left a comment. :) By the way, I think joriki's examples are already examples of what you're looking for. – 2011-07-27
-
0@Rajesh: Both of my examples can be restricted to the case $f: \mathbb{R} \to \mathbb{R}$. – 2011-07-27
2 Answers
Two examples come to mind -- the set of all functions normalized to $1$, and the set of Gaussians. (The latter is closed under multiplication -- don't know whether your "and" was intended to exclude that.)
[Edit in response to the comment:]
The product of two Gaussians is $\mathrm e^{a_1x^2+b_1x+c_1}\mathrm e^{a_2x^2+b_2x+c_2}=\mathrm e^{(a_1+a_2)x^2+(b_1+b_2)x+(c_1+c_2)}$, again a Gaussian.
By a function normalized to $1$, I mean a function whose sum/integral over its entire domain is $1$; for instance a probability distribution function.
[Edit in response to further comment:]
The space of functions normalized to $1$ is closed under convolution:
$$ \begin{eqnarray} \int_{-\infty}^\infty \left(\int_{-\infty}^\infty f(x)g(y-x)\mathrm dx\right)\mathrm dy &=& \int_{-\infty}^\infty f(x)\left(\int_{-\infty}^\infty g(y-x)\mathrm dy\right)\mathrm dx \\ &=& \int_{-\infty}^\infty f(x)\left(\int_{-\infty}^\infty g(y)\mathrm dy\right)\mathrm dx \\ &=& \int_{-\infty}^\infty f(x)\mathrm dx \\ &=& 1\;, \end{eqnarray} $$ and analogously for sums or multi-dimensional integrals.
-
0my 'and' is a logical one although i didn't have such an example in mind. But I do not understand how the set of Gaussians is closed under multiplications. Also please let me know what is meant by set of all functions normalized to 1. – 2011-07-27
-
0I didn't doubt that your 'and' was a logical one, but linguistically, since the repetition of "closed" and "under" was implied, it wasn't quite clear whether the repetition of "not" was also implied :-) – 2011-07-27
-
0The set of all functions normalized to 1 is not be closed under convolution unless you re-normalize it. – 2011-07-27
-
0just out of curiosity, do you think the examples given here are the only type of examples, or there could be more...and very different from these. – 2011-07-27
-
0@Rajesh: You can generalize Jyrki's example by starting with a whole set of functions and closing it under convolutions; if they're all centred on the same side of the origin, the resulting set won't be closed under addition. You can generalize my example to various other constraints on the total integral/sum $S$, for instance $0\le S\le c\le1$ or $|S|\le c\le1$. There might be further ones; I haven't thought about how one might characterize these sets systematically. – 2011-07-27
Convolution is associative, so pick you favorite function $f$ and start iterating. Define $f=f^{*[1]}$ and then recursive for all positive integers $n$ define $f^{*[n]}=f*(f^{*[n-1]})$. By associativity the set $$ S=\{f^{*[n]}\mid n\in\mathbf{Z}_+\} $$ is closed under convolution. If $f$ is not identically zero, then it cannot be closed under scalar multiplication by virtue of being a countable set. If we concentrate the mass of $f$ around 1 (for example use a Gaussian like joriki), then the mass of $f^{*[n]}$ will be concentrated (though not as sharply) around $n$, so set $S$ cannot be closed under multiplication or addition either.
Edit: This generalizes Theo Buehler's suggestion. Set $f=e_1$, and you get his example.
-
0Nice :-) ${}{}$ – 2011-07-27