Convergence lower bound method

Note that the nonlinearities involved in problem (33) are convex. Figure 11 shows the convergence of the OA and the GBD methods to the optimal solution using as a starting point y, = as = yy = 1. The optimal solution is Z = 3.5, with yi = 0. AS = 1. V3 = 0,. v, = 1,. V2 = 1. Note that the OA algorithm requires three major iterations, while GBD requires four, and that the lower bounds of OA are much stronger. 

Unfortunately there is no simple method of computing directly the joint spectral radius for two given general matrices. The best we can do is to compute the upper and lower bounds for larger and larger values of n and watch them converge. 

These two additional restrictions are implemented numerically. Identify two key independent design variables and provide realistic upper and lower bounds for these variables to assist the maximization algorithm in finding the best answer. The conjugate gradient optimization method should converge in approximately 20 iterations. 

Although the total energy calculated by DPT methods should in principle converge to the experimental value (-76.438 au), there are no upper or lower bounds for the currently employed methods with approximate exchange-correlation functionals. Indeed, all the gradient-corrected methods used here (BLYP, PBE and HCTH) give total energies well below the experimental value with the pc-4 basis set. 

One can use other general considerations to study the problem of polarizabilities and there are two common methods for doing so. One is based on the sum over eigenstates to the unperturbed Hamiltonian and is usually slowly converging because of the contributions from the continuum [6,14]. The other one is based on operator inequalities and can yield upper and lower bounds to the polarizabilities [llj. From the theory of operator inequalities [11,16] for the dipole polarizability in three dimensions we can write the ot2)zz component of the polarizability tensor in the form... 

In fact, it can be shown that Theorem 2 and its Corollary are vahd if the matrix T is replaced by the Gauss-Seidel matrix This gives a convergent method for obtaining upper and lower bounds for /I[oSPi], which is equal to by (4.9 ). 

The global optimization method aBB deterministically locates the global minimum solution of (1) based on the refinement of converging lower and upper bounds. The lower bounds are obtained by the solution of (15), which is formulated as a convex programming problem. Upper bounds are based on the solution of (1) using local minimization techniques. 

---