Consider the Duhamel's convolution $f*g$ of $f$ and $g$ defined by \begin{gather*} f*g(x)=\int_{0}^{x}f(x-\tau)g(\tau) d \tau. \end{gather*} Let $\hat{f}_c$ be the Fourier cosine transform of $f,$ that is, \begin{gather*} \hat{f}_c(\lambda)=\sqrt{\frac{2}{\pi}}\int_{0}^{+\infty}f(x)\cos(\lambda x) d x,\qquad x>0. \end{gather*} The inverse Fourier cosine transform $\check{g}$ of $g(\lambda)$ is essentially the same, which is defined as \begin{gather*} \check{g}_c(x)=\sqrt{\frac{2}{\pi}}\int_{0}^{+\infty}g(\lambda)\cos(\lambda x) d\lambda. \end{gather*} My question is: Is it true that \begin{gather*} \sqrt{\frac{2}{\pi}}f*g(x)=\left(\hat{f}_c\cdot\hat{g}_c\right)^{\check{~}}_c \end{gather*} for every $f$ and $g$ who behavior well enough?

I thought it is true, so I tried to prove it as follows. \begin{align*} &\big(f*g(t)\big)^{\hat{ }}_c=\sqrt{\frac{2}{\pi}}\int_{0}^{+\infty}\int_{0}^{t}f(t-\eta)g(\eta)d \eta \cdot \cos(t\lambda)d t\\ =& \sqrt{\frac{2}{\pi}}\int_{0}^{+\infty}\int_{0}^{+\infty}f(s)\cos(\lambda\eta+\lambda s) d s\cdot g(\eta)d \eta \\ =&\sqrt{\frac{2}{\pi}}\int_{0}^{+\infty} f(s)\cos(\lambda s)d s\cdot \int_{0}^{+\infty}g(\eta)\cos(\lambda \eta)d\eta\\ &-\sqrt{\frac{2}{\pi}}\int_{0}^{+\infty}\int_{0}^{+\infty}f(s)\sin(\lambda s)g(\eta)\sin(\lambda \eta)d\eta d s. \end{align*} Thus, if $\sqrt{\frac{2}{\pi}}\int_{0}^{+\infty}\int_{0}^{+\infty}f(s)\sin(\lambda s)g(\eta)\sin(\lambda \eta)d\eta d s=0,$ then the desired result follows. But is it true that $\sqrt{\frac{2}{\pi}}\int_{0}^{+\infty}\int_{0}^{+\infty}f(s)\sin(\lambda s)g(\eta)\sin(\lambda \eta)d\eta d s=0?$

PS: This question comes from the possible alternative solution of the Cauchy problem \begin{gather*}\tag{ODE} \begin{cases} w''(t)+\mu w(t)=f(t), \quad t>0,\\ w(0)=0, w'(0)=b, \end{cases} \end{gather*} where $\mu>0.$ I have solved this problem by using Duhamel's principle, and the method of variation of parameters. Thus, I want to try to solve by using Fourier cosine transform. I've finished calculating \begin{gather*} w(t)=\frac{b\sin(\sqrt{\mu}t)}{\sqrt{\mu}}+\left(\frac{1}{\mu-\lambda^2}\hat{ f}_c(\lambda)\right)^{\check{~ }}_c, \end{gather*} and the inverse Fourier cosine transform of $\frac{1}{\mu-\lambda^2}$ is \begin{gather*} \sqrt{\frac{\pi}{2\mu}}\sin(\sqrt{\mu}t). \end{gather*} Hence, if the answer to my question is yes, then the desired solution follows.

Add: After some thinking, my previous guess is wrong! Indeed, to solve (ODE), it suffices to construct a particular solution of the corresponding inhomogeneous ODE $w''(x)+\mu w(x)=f(x),$ by applying Fourier cosine transform. Apparently, this method is cumbersome, comparing to, for example, Laplace transformation, or Duhamel's principle.

  • 1
    $\begingroup$ Isn't the last integral the product of the sine Fourier transform of $f$ by the transform of $g$? $\endgroup$ Apr 28 '17 at 12:33
  • 1
    $\begingroup$ @HenriqueAugustoSouza, Not really. But it is, up to a constant. But this seems having nothing to do with my question. If $f$ is an even function on $\mathbb{R},$ then, in the derivation of Fourier cosine transform, I think the last integral is zero. $\endgroup$
    – azc
    Apr 28 '17 at 12:37
  • $\begingroup$ I see.. so you`re interested in the conditions for the last integral to vanish? Those who would make the equality true? $\endgroup$ Apr 28 '17 at 12:48
  • $\begingroup$ That property of the transform (exchanging products with convolutions) is expected from the full Fourier transform, so I believe it isn't true in general when restricted to it's sine and cosine summands. Sure, if just so happens that one of the summands is zero (such as when $f$ and $g$ are even, then the sine transform is zero), we get the equality. But I`m not sure that it's a necessary condition, we may have a weaker hypothesis. $\endgroup$ Apr 28 '17 at 12:54

The reason convolution works is that is a way of collecting objects of like kind. For example, $$ \sum_{n=0}^{\infty}a_n x^n\sum_{n=0}^{\infty}b_n x^n = \sum_{n=0}^{\infty}\left(\sum_{j+k=n}a_j b_k\right)x^n \\ \sum_{j+k=n}a_j b_k = \sum_{l=0}^{n}a_{l}b_{n-l} $$ Similarly, $$ \int_{0}^{\infty}e^{st}f(t)dt\int_{0}^{\infty}e^{st}g(t)dt= \int_{0}^{\infty}\left(\int_{0}^{t}f(u)g(t-u)du\right)e^{st}dt $$ Cosines are a little trickier. Let $\hat{f}$ denote the Fourier cosine transform. The question is: $$ \int_{0}^{\infty}\hat{f}(s)\cos(st)ds\int_{0}^{\infty}\hat{g}(s')\cos(s't)ds' = ? $$ Gathering like terms requires use of the identity $$ \cos(st)\cos(s't)=\frac{1}{2}\{\cos((s+s')t)+\cos((s-s')t)\}. $$ Collecting $s+s'=constant$ gives a convolution. The second terms require you to collect $|s-s'|=constant$, the absolute value coming from the fact that $cos(-a)=cos(a)$. So $$ ? = \frac{1}{2}\int_{0}^{\infty}\left(\int_{s'+s''=s}\hat{f}(s')\hat{g}(s'')+\int_{|s'-s''|=s}\hat{f}(s')\hat{g}(s'')\right)\cos(st)ds. $$ As before, $$ \int_{s'+s''=s}\hat{f}(s')\hat{g}(s'') = \int_{0}^{s}\hat{f}(s')\hat{g}(s-s')ds'. $$ And it would appear to me that $$ \int_{|s'-s''|=s}\hat{f}(s')\hat{g}(s'') = \int_{s'-s''=s}\hat{f}(s')\hat{g}(s'')+\int_{s''-s'=s}\hat{f}(s')\hat{g}(s'') \\ = \int_{0}^{\infty}\hat{f}(s''+s)\hat{g}(s'')ds''+\int_{0}^{\infty}\hat{f}(s')\hat{g}(s'+s)ds' $$ So my guess is that convolution is replaced by $$ \hat{f}\star_{c}\hat{g} = \frac{1}{2}\left[\int_{0}^{s}\hat{f}(s')\hat{g}(s-s')ds'+\int_{0}^{\infty}\hat{f}(s''+s)\hat{g}(s'')ds''+\int_{0}^{\infty}\hat{f}(s')\hat{g}(s'+s)ds'\right] $$ Then there is reason to suspect that $$ f(t)g(t) = \int_{0}^{\infty}(\hat{f}\star_{c}\hat{g})(s)\cos(st)ds. $$ But that's just my guess.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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