# Is convolution operator compact?

I know convolution is not a Hilbert–Schmidt integral operator, but it needs more to tell if convolution is compact or not.

-
What properties does the function being convoluted with have? Between what spaces are you considering this operator? – user12014 Jul 11 '12 at 21:53
Let's say it's in Hilbert space. – chaohuang Jul 12 '12 at 0:03
You mean $L^2(\mathbb{R}^n)$, I guess. But what is the precise definition of your convolution operator? – Giuseppe Negro Jul 12 '12 at 0:36

Let us use $T_f$ to denote the operator which convolutes with $f$ and let $M_\phi$ to denote the operator which multiplies by $\phi$. Since Young's inequality provides sharp bounds, we necessarily have $f \in L^1$, if we want $T_f:L^2\to L^2$ to even be bounded. Consider $FT_f$, the composition of the Fourier transform with $T_f$. Since $F$ is unitary, $T_f$ is compact if and only if $FT_f$ is. By the convolution theorem we have

$$FT_fg = \hat{f}\cdot\hat{g} = M_{\hat{f}}Fg$$

so that $FT_f = M_{\hat{f}}F$ is a multiplication operator composed with the Fourier transform. Again, the operator on the right hand side is compact if and only if the multiplication operator $M_{\hat{f}}$ is. Now, $\hat{f} \in C_0$ and hence $\hat{f} \in L^\infty$. But the only compact multiplication operator on $L^2$ induced by a bounded measurable function is the operator that is identically zero. Hence $\hat{f}$ and therefore $f$ must be identically zero. It follow that the only compact convolution operator on $L^2$ is the operator which is identically zero.

To see that there are no non-trivial compact multiplication operators, suppose that $f \in L^\infty$ and $f \neq 0$. Then there exists a set $E$ of positive measure such that $|f| \geq \epsilon > 0$ on $E$. Thus $M_f$ would be compact with a bounded left inverse on a subspace isomorphic to $L^2(E)$ which is impossible.

-
You can also notice that the spectrum of $T_f$ contains $\{\hat{f}(t),t\in\mathbb{R}\}$, which is not countable – San-A Mar 22 '13 at 11:01

good questions but generally speaking, never. Over on the fourier transform side divide the set where $\hat f > \epsilon$ into infinitely many disjoint intervals then the inverse transforms of the intervals, properly normalized, map to infinitely many orthogonal guys that don't go to zero

-
I downvoted. This answer means nothing, you should be more clear. – Giuseppe Negro Jul 11 '12 at 22:57
sorry, i see your clarification . answer is especially clear in $\mathbb L^2$ – mike Jul 12 '12 at 0:06
"the inverse transforms of the intervals". What does inverse transform of intervals mean?@mike – Harish Feb 6 '15 at 16:28