Estimations Hessians

Figure A.2 Finite-difference estimated Hessian and variance (a = 10, /3 = 0.1)...

In simple relaxation (the fixed approximate Hessian method), the step does not depend on the iteration history. More sophisticated optimization teclmiques use infonnation gathered during previous steps to improve the estimate of the minunizer, usually by invoking a quadratic model of the energy surface. These methods can be divided into two classes variable metric methods and interpolation methods. 

Schlegel H B 1984 Estimating the Hessian for gradient-type geometry optimizations Theor. Chim. Acta 66 333... 

Most optimization algorithms also estimate or compute the value of the second derivative of the energy with respect to the molecular coordinates, updating the matrix of force constants (known as the Hessian). These force constants specify the curvature of the surface at that point, which provides additional information useful for determining the next step. 

A more sophisticated version of the sequential univariate search, the Fletcher-Powell, is actually a derivative method where elements of the gradient vector g and the Hessian matrix H are estimated numerically. 

Almost all optimization methods need a starting geometry and an initial estimate of the Hessian. The Hessian is improved as the optimization proceeds. 

To determine the full set of normal modes in a DFT calculation, the main task is to calculate the elements of the Hessian matrix. Just as we did for CO in one dimension, the second derivatives that appear in the Hessian matrix can be estimated using finite-difference approximations. For example,... 

For a description of the NEB method and a comparison of the NEB method with other chain of states methods for determining transition states without the use of the Hessian matrix, see D. Sheppard, R. Terrell, and G. Henkelman, J. Chem. Phys. 128 (2008), 134106. The climbing image NEB method is described in G. Henkelman, B. P. Ubemaga, and H. Jonsson, J. Chem. Phys. 113 (2000), 9901 9904. Improvements to the NEB method including a better estimation of the tangent to the MEP are described in G. Henkelman and H. Jonsson,... 

Vibrational frequencies measured in IR experiments can be used as a probe of the metal—ligand bond strength and hence for the variation of the electronic structure due to metal—radical interactions. Theoretical estimations of the frequencies are obtained from the molecular Hessian, which can be straightforwardly calculated after a successful geometry optimization. Pure density functionals usually give accurate vibrational frequencies due to an error cancellation resulting from the neglect of... 

There are a few points with respect to this procedure that merit discussion. First, there is the Hessian matrix. With elements, where n is the number of coordinates in the molecular geometry vector, it can grow somewhat expensive to construct this matrix at every step even for functions, like those used in most force fields, that have fairly simple analytical expressions for their second derivatives. Moreover, the matrix must be inverted at every step, and matrix inversion formally scales as where n is the dimensionality of the matrix. Thus, for purposes of efficiency (or in cases where analytic second derivatives are simply not available) approximate Hessian matrices are often used in the optimization process - after aU, the truncation of the Taylor expansion renders the Newton-Raphson method intrinsically approximate. As an optimization progresses, second derivatives can be estimated reasonably well from finite differences in the analytic first derivatives over the last few steps. For the first step, however, this is not an option, and one typically either accepts the cost of computing an initial Hessian analytically for the level of theory in use, or one employs a Hessian obtained at a less expensive level of theory, when such levels are available (which is typically not the case for force fields). To speed up slowly convergent optimizations, it is often helpful to compute an analytic Hessian every few steps and replace the approximate one in use up to that point. For really tricky cases (e.g., where the PES is fairly flat in many directions) one is occasionally forced to compute an analytic Hessian for every step. 

The results of estimation of the system by direct maximum likelihood are shown. The convergence criterion is the value of Belsley (discussed near the end of Section 5.5). The value a shown below is g H" g where g is the gradient and H is the Hessian of the log-likelihood. 

If we had estimates in hand, the simplest way to estimate the expected values of the Hessian would be to evaluate the expressions above at the maximum likelihood estimates, then compute the negative inverse. First, since the expected value of <51nL/3a is zero, it follows that E[x/5] = 1/a. Now,... 

Notice that in spite of the quite different coefficients, these are identical to the results for the probit model. Remember that we originally estimated the probabilities, not the parameters, and these were independent of the distribution. Then, the Hessian is computed in the same manner as for the probit model using hy = Fi/l-Fy) instead of X0Xx in each cell. The asymptotic covariance matrix is the inverse of... 

There is more than one way to estimate the parameters. As in Example 5.13, the method of scoring (using the expected Hessian) will be straightforward in principle - though in our example, it does not work well in practice, so we use Newton s method instead. The iteration, in which we use index to indicate the estimate at iteration t, will be... 

The zero off diagonal elements in the expected Hessian make this convenient, as the iteration may be broken into two parts. We take the iteration for u first. With current estimates u(7) and y(/), the method of... 

Global strategies for minimization are needed whenever the current estimate of the minimizer is so far from x that the local model is not a good approximation to fix) in the neighborhood of x. Three methods are considered in this section the quadratic model with line search, trust region (restricted second-order) minimization and rational function (augmented Hessian) minimization. 

---