The HFB state is a linear combination of eigenstates of the particle-number operator, i.e.,
(81) |
The HFB+VAPNP variational procedure gives, in principle, the same value of the projected energy independently of the value of . This independence can, however, be subject to numerical instabilities whenever the amplitude , corresponding to the projection on particles, is small. Therefore, for practical reasons, one is interested in keeping the average number of particles as close as possible to , which guarantees that amplitude is as large as possible.
In the standard HFB equations (21), the average number of particles is kept equal to by adjusting the Lagrange multiplier . However, in the HFB+VAPNP approach, does not appear in the variational equations (65), because the variation of the constant term equals zero. Therefore, the HFB+VAPNP equations (65) do not allow for adjusting the average particle number , which, during the iteration procedure, may become vary different from . Moreover, such uncontrolled changes of from one iteration to another may preclude reaching the stable self-consistent solution.
In order to cope with these problems, one can artificially reintroduce a constant , analogous to the Fermi energy , into the HFB+VAPNP equations (65), i.e.,
Had such an ideal situation continued till the end, a nonzero value of would have never appeared, and the required solution would have been found. In practice, this situation never happens, and at some iteration one finds that Tr, i.e., the sum of norms of the second components (68) of the HFB+VAPNP wave functions, is larger (smaller) than . In such a case, in the next iteration one uses a slightly negative (positive) value of , which decreases (increases) the norms of , and decreases (increases) the average particle number in the next iteration. Since acts in exactly the same way as the Fermi energy does within the standard HFB method, the well-established algorithms of readjusting can be used. Moreover, as soon as the iteration procedure starts to converge, the non-zero values of cease to be needed, and thus naturally converges to zero, as required. In practice, we find that the above algorithm is very useful, and it provides the same converged solution with any value of , being a small integer.