However, how can we be sure that the optimized geometry given by the calculation really is an energy minimum? 

 

It could be a different type of stationary point (a saddlepoint). 

 

Remember a (first-order) saddlepoint is a maximum in one direction, and a minimum in all other directions. 

 

For a simple function in one dimension, at a minimum the second derivative (f'') is positive whereas at a maximum it is negative. 

 

For a multidimensional function, we need to calculate the eigenvalues of the Hessian matrix. 

 

The eigenvalues, l, are given by:

(H lI)x = 0

x is an eigenvector, which we will deal with later. 

 

For a non-trivial solution, the determinant must be zero, i.e.

 

For large matrices, this is done by diagonalizing the matrix. 

 

This involves finding the matrix U so that

U-1HU = E

where E is the matrix of eigenvalues. 

 

Standard computational methods have been developed for diagonalizing matrices, which we do not need to go into. 

 

There are 3N eigenvalues (N being the number of atoms) if Cartesian coordinates are used. 

 

Next: using the eigenvalues