1
$\begingroup$

Okay, some terminology, correct me if I'm wrong.

Singular support - the set on which a distribution fails to be smooth. In this case a piecewise function.

Is there a name for $f*f*f$? The convolution of function with itself $f*f$ is called the autocorrelation, right.

I want to know the singular support of $f*f*f$, where $f$ is the indicator function of a regular hexagon in the plane?

I can calculate, in mathematica, $f*f$ - it's a Piecewise function. I can extract the conditions without inequalities (ie line segments) and plot it. (For some reason it seems to include the axes, on which I know $f*f$ is smooth).

singular support of f*f

However, trying this for $f*f*f$ is overheating my Raspberry Pi.

Also, I know from the Titchmarsh Convolution Thereom, that since supp f is convex. $$ supp f*f*f \subset supp f +supp f*f \subset supp f+supp f +supp f, $$ it seems that $supp f*f*f = 3 supp f$.

Is there a version of the Titchmarsh Convolution theorem for singular support?

$\endgroup$

2 Answers 2

1
$\begingroup$

The answer is all line segments of the form $p_1 + p_2 + e$ where $p_1$ and $p_2$ are vertices of the hexagon and $e$ is an edge. More carefully, I can prove that the singular locus is contained in this set, and I expect equality. Here is a picture.

enter image description here

Proof: Write $H$ for the filled in hexagon. So $H^3$ is a $6$-dimensional polytope in $\mathbb{R}^6$. There is a linear map $\pi: H^3 \to \mathbb{R^2}$ sending $(x_1, y_1, x_2, y_2, x_3, y_3) \to (x_1+x_2+x_3, y_1+y_2+y_3)$. The (ordinary) support of $f \ast f \ast f$ is the image of $\pi$, and the values of the convolution are the $4$-dimensional volumes of the fibers of $\pi$.

Now, as $z$ varies through $\mathbb{R}^2$, $\pi^{-1}(z)$ is a four dimensional polytope. Generically, changing $z$ will just slide the defining facets of $\pi^{-1}(z)$ parallel to themselves. The volume of a polytope with fixed facet normals is a polynomial (degree $4$ in this case) in the positions of those facets, hence smooth.

When do singularities occur? When the combinatorics of $\pi^{-1}(z)$ changes. This happens when $5$ facets of $\pi^{-1}(z)$ pass through a common vertex. (Since these are $5$ hyperplanes in $4$-space, it doesn't happen generically.) Back in the big $6$ dimensional polytope $H^3$, this is the same as saying that $z$ lies below a point of $H^3$ which is on $5$ facets -- in other words, $z$ lies below an edge of $H^3$. That's the condition I have above.

Now, if $z$ only lay below one edge of $H^3$, I could write down the singularity introduced by that edge and show that $f \ast f \ast f$ is indeed singular there. Instead, most points which are below some edge of $H^3$ are below many edges, so I can't guarantee the singularities don't cancel. But I don't expect them to.


To be more general, let $P$ be a full dimensional subpolytope in $\mathbb{R}^d$ and let $\pi: \mathbb{R}^d \to \mathbb{R}^e$ be a linear map. Let $g(z) = \mathrm{Volume}(\pi^{-1}(z))$ for $z \in \mathbb{R}^e$. Then $g$ is singular along the images of the $e-1$ dimensional faces of $P$.

If $f_1$, $f_2$, ..., $f_r$ are the characteristic functions of polytopes $P_1$, $P_2$, ..., $P_r$ in $\mathbb{R}^e$, then $f_1 \ast f_2 \ast \cdots \ast f_r$ is singular along sets of the form $Q_1 + Q_2 + \cdots + Q_r$, where $Q_j$ is a face of $P_j$ and $\sum \dim Q_j = e-1$

$\endgroup$
6
  • 2
    $\begingroup$ One can also use the theory of distributions. The third (distributional) derivative $\nabla^3(f*f*f)$ of $f$ is equal to $\nabla f * \nabla f * \nabla f$ (distributing indices appropriately). $\nabla f$ is supported (again, as a distribution) on the boundary of the hexagon, which is compact, hence $\nabla f * \nabla f * \nabla f$ is supported on the Minkowski sum of three copies of this boundary, and so these are the only places where $f*f*f$ can be singular. $\endgroup$
    – Terry Tao
    Aug 22, 2015 at 20:33
  • $\begingroup$ Great, thanks for that. BTW these are "hexsplines", see bigwww.epfl.ch/demo/hexsplines. The argument you provided seems similar to the construction of box-splines. $\endgroup$
    – pdmclean
    Aug 22, 2015 at 21:52
  • $\begingroup$ @TerryTao But that gives you a much weaker statement, unless I am confused. Looking at the case of $f \ast f$, the Minkoski sum of two copies of $\partial H$ is the entire interior of $2H$, not the $1$-dimensional thing we want. $\endgroup$ Aug 22, 2015 at 22:00
  • $\begingroup$ I'll try to confirm this computationally somehow. $\endgroup$
    – pdmclean
    Aug 22, 2015 at 22:01
  • $\begingroup$ Gah, you're right. One needs the additional observation that the convolution of two distributions that are one-dimensionally smooth on non-parallel line segments will be two-dimensionally smooth except on the boundary of the supporting parallelogram, and upon taking another derivative one gets smooth measures on the one-dimensional boundary edges plus some zero-dimensional stuff on the vertices. This should eventually let one arrive at the same 1-skeleton that you have in your answer. Probably the "right" way to proceed is via wave front sets, though... $\endgroup$
    – Terry Tao
    Aug 22, 2015 at 22:24
-1
$\begingroup$

supp$f*f*f$ is just $H+H+H$ where $H$ is the hexagon. If $C$ is any convex set, then $C+C+\ldots+C$ ($n$ times) is just $nC=\{nx\colon x\in C\}$.

To see this, notice that if $x_1,\ldots,x_n\in C$, then $x_1+\ldots+x_n=n\cdot \frac1n(x_1+\ldots+x_n)\in nC$. Conversely if $x\in nC$, then $x=ny$ for $y\in C$, so that $x=y+y+\ldots+y$, so that $x\in C+\ldots+C$.

$\endgroup$
2
  • $\begingroup$ Nice to see how that bit works. Thanks. $\endgroup$
    – pdmclean
    Aug 22, 2015 at 2:32
  • $\begingroup$ I am a bit puzzled. The OP asked about the singular support, not the usual one. For instance, if $f$ is the indicator function of the unit interval, then the singular support of $f*f*f$ is the set $\{0,1,2,3\}$ rather than the interval $[0,3]$. Тhe picture above apparently represents the singular support of $f*f$ (although I don't understand the meaning of colors). Now, I do not see anything in the current answer which would indicate that $f*f*f$ is nowhere smooth on its ordinary support. $\endgroup$
    – R W
    Aug 22, 2015 at 14:24

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.