A variational method for numerical differentiation (Q1347030)

A method is produced for effecting numerical differentiation by means of an optimization procedure. Given a real valued, smooth function \(u\) on \([a,b]\), an associated functional \[ H(q) = \int_ a^ b({u'}^ 2 + qu^ 2)dx - \int_ a^ b({u_ q'}^ 2 + qu_ q^ 2)dx \] is constructed, where \(u_ q\) solves the boundary value problem \[ -v'' + qv = 0 \qquad v(a) = u(a),\;v(b) = u(b) . \] Its properties are studied; in particular, \(H\) is strictly convex and attains its unique global minimum at \(Q = u''/u\). The numerical computation of \(Q\) makes use of a steepest descent procedure. Once an approximation of \(Q\) is found, \(u'\) is estimated numerically solving \(-u'' + Qu = 0\) as a first-order system under appropriate boundary conditions. The solvability and convergence are discussed; some numerical examples are presented and results are compared with other methods. Finally, a higher-order version of the functional \(H\) is considered.