Maximum entropy solution to ill-posed inverse problems with approximately known operator (Q2427763)

The authors want to estimate an unknown measure \(\mu(x)\) from the data \(y^{obs}=y+\varepsilon\), where \(y=\int_X\Phi(x)\,d\mu(x)\) and \(\varepsilon\) is noise, \(\|\varepsilon\|<\eta\), \(\eta>0\) is the ``noise level'', \(X\subset \mathbb{R}^d\), \(\Phi:\mathbb{R}^d\to\mathbb{R}^k\) is a continuous map, known approximately, and \(\mu\) is an unknown finite measure. They propose an ``approximate maximum entropy on the mean (AMEM)'' estimate for \(\mu\), given \(y^{obs}\), and claim the (weak) convergence of the estimate. Reviewer's remarks. There are unclear points in this paper. For example, assume for simplicity that \(d = k = 1\), \(\Phi(x) = x\) is known exactly, and \(\varepsilon=0\). Then their problem is: given \(y =\int_X xd\mu(x)\), estimate \(\mu(x)\). This problem has infinitely many solutions. Selecting a particular solution can be done in infinitely many ways. The problem itself is of limited theoretical and practical interest. Moreover, since the weak convergence is claimed for a fixed noise level (\(\eta>0\)), the usefulness of this convergence is not clear.