Symmetry of dispersion/transport for convolutions with real-symmetric kernels

55 Views Asked by At

I am new to math-exchange, so please let me know if I state my question in an incomprehensible way. My problem comes from the field of wave optics but can be stated in a quite general setting:

Problem statement: Given a real-valued function $f \in L^2(\mathbb R)$, that vanishes on the negative half-axis, $f|_{\mathbb R_ -} = 0$, consider the convolution $S(f):= s \ast f$ with a real-valued and symmetric kernel $s \in L^1(\mathbb R)$, i.e. $s(x) = s(-x) \in \mathbb R$ for all $x \in \mathbb R$. Is there a general bound of the form

$\|S(f)|_{\mathbb R_-}\|_{L^2} \leq C \|S(f)|_{\mathbb R_+}\|_{L^2}$

for some constant $C \geq 1$? It can be readily seen that this cannot hold for any $C < 1$ but numerical experiments indicate that it might hold for $C = 1$.

The idea behind it: Intuitively, the idea behind my question is whether convolutions with symmetric kernels also transport the mass of $f$ in some sufficiently symmetric manner into the negative- and positive direction.

Preliminary thoughts: I tried to take advantage of the fact that the operator $S$ commutes with the (anti-)symmetrization operations, $g^{\text s} := \frac 1 2 (g + g^-)$ and $g^{\text a} := \frac 1 2 (g - g^-)$ where $g^-(x) := g(-x)$:

$S(g)^{\text s} = S(g^{\text s})$ and $S(g)^{\text a} = S(g^{\text a})$.

Moreover, I think that it may be of use that the decomposition is orthogonal: $\|g\|_{L^2}^2 = \|g^{\text s}\|_{L^2}^2 + \|g^{\text a}\|_{L^2}^2$ for all real-valued $g \in L^2(\mathbb R)$. Finally, the condition $f|_{\mathbb R_-} = 0$ can be equivalently stated as

$f^{\text a}|_{\mathbb R_+} = f^{\text s}|_{\mathbb R_+}$ and $f^{\text a}|_{\mathbb R_-} = - f^{\text s}|_{\mathbb R_-}$.

Summary: I have the feeling that the problem is so general that there could already be an exhaustive and rather well-known answer to it. If so, however, I really do not know where to find it, probably lacking the right buzzwords.

Thanks for help!

1

There are 1 best solutions below

0
On

The answer is NO. This was pointed out to me be Cong Shi in personal communication.

Counter-example: For simplicity, we take $s(x):= \delta( x + \frac 1 2 ) + \delta ( x - \frac 1 2 )$, where $\delta$ denotes the Dirac-delta-distribution. Then $s$ is real-valued and symmetric but technically $s \notin L^1(\mathbb R)$. However, we can approximate $s$ by a suitable Dirac-sequence $(s_k)_{k \in \mathbb R} \subset L^1(\mathbb R)$ to make the constructed example apply rigorously to the given setting.

Now, for $0 < a < 1$ and $0 < \varepsilon < \frac 1 2$, we define

$f(x) := \sum_{k=0}^\infty (-a)^k \boldsymbol 1_{[0; \varepsilon]} ( x - k), \qquad \boldsymbol 1_{[0; \varepsilon]}(x):= \begin{cases} 1 &\text{if }x \in [0; \varepsilon] \\ 0 &\text{else} \end{cases}$.

Then the convolution with $s$ can be computed analytically:

$(s \ast f)(x) = \sum_{k=0}^\infty (-a)^k \boldsymbol 1_{[0; \varepsilon]} ( x - k + \frac 1 2) + \sum_{k=0}^\infty (-a)^k \boldsymbol 1_{[0; \varepsilon]} ( x - k - \frac 1 2) \\ = \boldsymbol 1_{[0; \varepsilon]} ( x + \frac 1 2) + \sum_{k=0}^\infty ((-a)^k + (-a)^{k+1}) \boldsymbol 1_{[0; \varepsilon]} ( x - k - \frac 1 2) \\ = \boldsymbol 1_{[0; \varepsilon]} ( x + \frac 1 2) + (1-a) \sum_{k=0}^\infty (-a)^k \boldsymbol 1_{[0; \varepsilon]} ( x - k - \frac 1 2)$.

The first and second summand in the bottom line correspond to the parts of $s \ast f$ that are supported in the negative- and positive half axis, respectively. Accordingly, we obtain

$\|S(f)|_{\mathbb R _-}\|_{L^2}^2 = \|(s\ast f)|_{\mathbb R _-}\|_{L^2}^2 = \int_{\mathbb R_-} \boldsymbol 1_{[0; \varepsilon]} ( x + \frac 1 2) ^2 \,\text d x = \int_{-\frac 1 2}^{\varepsilon-\frac 1 2} 1 \, \text d x = \varepsilon \\ \|S(f)|_{\mathbb R _+}\|_{L^2}^2 = \|(s\ast f)|_{\mathbb R _+}\|_{L^2}^2 = (1-a)^2 \sum_{k=0}^\infty a^{2k} \int_{\mathbb R_+} \boldsymbol 1_{[0; \varepsilon]} ( x - k - \frac 1 2)^2 \, \text d x \\ = (1-a)^2 \sum_{k=0}^\infty a^{2k} \int_{k + \frac 1 2}^{k + \frac 1 2 +\varepsilon} 1 \, \text d x \\ = (1-a)^2 \varepsilon \sum_{k=0}^\infty a^{2k} = \varepsilon \cdot \frac{(1-a)^2}{1-a^2} = \varepsilon \frac{ 1-a }{1+a}$.

The result implies that $\|(s\ast f)|_{\mathbb R _+}\|^2/\|(s\ast f)|_{\mathbb R _-}\|^2 = \frac{ 1-a }{1+a}$ is always less than 1 and even goes to 0 for $a \to 1$. Hence, there cannot exist any constant $C > 0$ such that

$\|S(f)|_{\mathbb R _-}\|_{L^2} \leq C \|S(f)|_{\mathbb R _-}\|_{L^2}$

holds true in the general problem setting.