For the equality-constrained minimization problem, and . Put . The restricted objective isThe only zero has , so . Since , the second-derivative criterion for convexity proves a unique minimizing . Consequently the equality constraint has minimizers and , with minimum .
For the inequality-constrained minimization problem, the identity supplies the universal lower bound . Both equality points for that bound, and , satisfy . Hence the inequality constraint has exactly these two minimizers, with minimum .
The stationary point equations give and , hence . At the Hessian matrix has negative determinant , so is a saddle point. At each of and , the Hessian matrix has positive leading minors and , proving a strict local minimum. More strongly,Thus and are exactly the global minima, with value . There are no local maxima, since any differentiable interior local maximum would be a stationary point, and all three have been classified.
The derivative is represented by the gradient, and the Hessian matrix contains the second partial derivatives:Equivalently the derivative as a linear map is .
Write the computed function values as , where . The computed central finite difference satisfiesThe two contributions are respectively evaluation absolute error and truncation error. For , step-size choice for a noisy central derivative minimizes their sum: its derivative is , negative before and positive afterIf , the bound tends to zero as and has no positive minimizing step.
Apply the preceding Taylor's theorem at and , with . The even-order terms cancel in the central finite difference, givingBoth Lagrange remainders have magnitude at most , so the triangle inequality gives the sharp central-secant derivative error
By Taylor's theorem with Lagrange remainder, for an increment ,where lies between and . The given third derivative bound implies .
There are unlisted articles, also show them or only show them.