The Sylvester and Bézout resultant matrices for blind image deconvolution (Q1799667)

This paper compares the use of the Sylvester and Bézout matrices for the solution of the problem of Blind Image Deconvolution. This problem is solved by computing an approximate gcd (AGCD) of two bivarialte polynomials, and actually, the discussion in the paper focuses on the calculation of an AGCD of two univariate polynomials. The similarities and the differences between both matrices are well known. Likewise, countless methods are known to calculate a gcd from both matrices. When calculating an agcd, the real problem lies in the calculation of its degree, which is equal to the rank loss of both matrices. Having said that, the authors compare two methods, one with the Sylvester matrix, and the other with the Bézout matrix. They claim that the Sylvester matrix yields betther results because it returns the correct degree of the AGCD, and coefficients with a small backward error, for much higher noise levels than the Bézout matrix. Moreover, computational examples of blurred and deblurred images show the superior results obtained with the Sylvester matrix. However, as the reader reads the article, it is appreciated a lack of precision and rigor in describing the method with the Bézout matrix. For example, as conclusion, the use of the leading principal submatrices (Theorem 2) of the Bézout matrix for computing the degree is not recommended because they prone to numerical instability. But this fact is well known because the leading principal minors of Bezoutians are the principal subresultants (up to a constant), and the Subresultant Chain of two polynomials can be defective (see Example 2). Curiously, this is not mentioned along the paper.