3
$\begingroup$

My background is chemistry and I was exploring some one dimensional deconvolution problems i.e., resolution of two or more overlapping peaks. A lot of excellent work was done in the 1970-80s. However, I wanted to clarify a conceptual point or proper mathematical nomenclature since most of these signal deconvolution problems are not written by mathematicians. Starting from convolution integral, the typical starting discussion is:

$$ C(E)=\int_{-\infty}^{\infty} W\left(E^{\prime}\right) R\left(E-E^{\prime}\right) \mathrm{d} E^{\prime} $$ with $$ \int_{-\infty}^{\infty} R(E) \mathrm{d} E=1 $$

where $E$ and $E^{\prime}$ denote the radiation energy; $C(E)$, $E(E)$ and $R(E)$ mean the observed spectrum, the intrinsic distribution and the response function, respectively. The goal is to obtain true distribution is obtained from deconvoluting the observed spectrum with the response function.

Most authors then switch to the following matrix equation, $$ \mathrm{C}=R W $$ where $R$ is a square matrix, and $C$ and $W$ are column vectors which correspond to the observed spectrum after background subtraction and to the intrinsic distribution, respectively.

(i) My understanding is that convolution by matrices is a circular convolution and this is done by circulant matrices. Is there a matrix formulation that is meant for linear convolution, or zero padding is the only way to obtain linear convolution out of circulant matrices? I mean is there a special matrix, besides circulant matrix, which performs linear convolution?

(ii) Sometimes the name Toepltiz matrix is used in deconvolution problems. Should we call $R$ a circulant matrix or a Toeplitz matrix?

I would appreciate a readible reference on this topic.

$\endgroup$
7
  • 1
    $\begingroup$ Any linear operation on a finite-dimensional based vector space is performed by a matrix. But why do these signals live in a finite-dimensional space? $\endgroup$
    – LSpice
    Commented Nov 19, 2021 at 4:06
  • $\begingroup$ 1) a Toeplitz matrix (constant diagonals) can be used to represent a linear convolution, see for example these notes; 2) a circulant matrix is a special type of Toeplitz matrix; 3) don't know. $\endgroup$ Commented Nov 19, 2021 at 9:22
  • $\begingroup$ @CarloBeenakker, Thanks for the useful link. In the notes they also zero padded the matrix to make it a linear convolution. $\endgroup$
    – ACR
    Commented Nov 19, 2021 at 13:05
  • $\begingroup$ @LSpice, As to your point "But why do these signals live in a finite-dimensional space?" I think this is just a utilitarian approach. A real experiment has a starting time and an end time. Consider, collecting an infrared spectrum of water. A given infrared spectrometer will have a wavelength measurement range, again finite data. $\endgroup$
    – ACR
    Commented Nov 19, 2021 at 13:07
  • $\begingroup$ In light of considering finite-dimensional signals, what do you mean by linear convolution? It seems to me that this is nothing else but zero-padding. That is, I don't know how to perform a linear convolution without regarding signals as extended by $0$ as necessary, and then the matrix implementing the convolution will have $0$s in the appropriate places. $\endgroup$
    – LSpice
    Commented Nov 19, 2021 at 13:24

0

You must log in to answer this question.