Optimal finite difference grids for direct and inverse Sturm$ndash$Liouville problems (Q3146255)

The authors consider the construction of `optimal' finite difference grids for Sturm-Liouville problems in the isotropic form \[ {d\over dx}\left(\sigma(x){du\over dx}\right) = \lambda \sigma(x)u(x), \quad x\in (0,L), \] with regular endpoints. The function \(\sigma\) is positive. The grids constructed are `optimal' in the sense that they give optimal approximations to the Dirichlet to Neumann map \(f(\lambda)\), also known in the ODE case as the Titchmarsh-Weyl function. In fact, the paper contains rather more information than the title suggests. Firstly, there is an extremely useful review of the theory of inverse eigenvalue problems for three-term finite difference schemes. This goes some way to explaining the paradox that continuous inverse Sturm-Liouville problems seem, a priori, to require twice as much spectral data (two spectra corresponding to different boundary conditions) as their finite difference analogues. The explanation lies in the fact that one must not regard the meshpoints as known. Rather, there are two meshes to be recovered, called \((z_j)\) and \((\hat{z}_j)\); both these meshes play a rôle in the construction of the finite difference scheme, with the coefficient values at both sets of meshpoints appearing in different parts of the scheme \[ {1\over \hat{h}_j\sigma_j}\left\{\hat{\sigma}_j{U_{j+1}-U_j \over h_j} - \hat{\sigma}_{j-1}{U_{j}-U_{j-1}\over h_{j-1}}\right\} = \lambda U_j, \] with \(h_j = z_j-z_{j-1}\), \(h_{j-1}=\hat{z}_{j}-\hat{z}_{j-1}\), \(\sigma_j = \sigma(z_j)\), \(\hat{\sigma}_j = \sigma(\hat{z}_j)\). The finite difference scheme clearly depends precisely on the sequences of \(\hat{h}_j\sigma_j\) and \(h_j/\hat{\sigma}_j\), and it is precisely these two sequences which the theory of discrete inverse problems allows to be determined. At this point, the paradox seems to be reversed. The discrete problem seems to require more information than the continuous one: from the quantities \(\hat{h}_j\sigma_j\) and \(h_j/\hat{\sigma}_j\) one cannot uniquely determine approximations to the coefficients. One would need to know the \(\hat{h}_j\) and \(h_j\). So, in fact, the second important aspect of the paper is that the authors give constructive formulae for determining the \(\hat{h}_j\) and \(h_j\) from a known function \(\sigma_0(\cdot)\) which have the rather desirable property that when used to solve an inverse problem for an unknown \(\sigma(\cdot)\), they give convergent results. The convergence is in a rather weak norm (see equation (4.40)) but this is the best that one can expect for inverse problems. A very important point here is that the simple choice of a uniform mesh, \(h_j = \hat{h}_j = h\) (constant) with \(z_j = jh\) and \(\hat{z}_{j+1/2} = (j+1/2)h\) does not work. However, meshes constructed from a constant function \(\sigma_0(\cdot)\) do work. The error analysis is in section 5 and proceeds by reducing the problems to Liouville normal form, where classical asymptotics in the spectral parameter can be used. The main contribution to the error from a given mesh, for large \(| \lambda| \), is shown to come from the Fourier case \(y'' + \lambda y = 0\). This result should be compared with the work of Andrew, who has shown that, in order to solve an inverse Sturm-Liouville problem by finite differences on a uniform grid, it is necessary first to perturb the eigenvalues by adding to them the errors which would arise if they were approximated by finite differences. These errors can be approximated by considering the Fourier case.