SOFTSUSY
4.1
|
Contains two-loop routines from Slavich et al. More...
Go to the source code of this file.
Functions | |
int | ewsb2loop_ (double *t, double *g, double *t1, double *t2, double *st, double *ct, double *q, double *mu, double *tanb, double *vv, double *gs, double *s1, double *s2) |
int | dszhiggs_ (double *t, double *g, double *T1, double *T2, double *st, double *ct, double *q, double *mu, double *tanb, double *v2, double *gs, int *OS, double *S11, double *S22, double *S12) |
int | dszodd_ (double *rmtsq, double *mgsq, double *mst1sq, double *mst2sq, double *sxt, double *cxt, double *scalesq, double *amu, double *tbeta, double *vev2, double *gstrong, double *p2s) |
int | ddshiggs_ (double *t, double *b, double *A0, double *T1, double *T2, double *B1, double *B2, double *st, double *ct, double *sb, double *cb, double *q, double *mu, double *tanb, double *vv, double *S11, double *S12, double *S22) |
int | ddsodd_ (double *t, double *b, double *A0, double *T1, double *T2, double *B1, double *B2, double *st, double *ct, double *sb, double *cb, double *q, double *mu, double *tanb, double *vv, double *dMA) |
int | ddstad_ (double *t, double *b, double *A0, double *T1, double *T2, double *B1, double *B2, double *st, double *ct, double *sb, double *cb, double *q, double *mu, double *tanb, double *vv, double *s1, double *s2) |
int | tausqodd_ (double *t, double *A0, double *BL, double *T1, double *T2, double *st, double *ct, double *q, double *mu, double *tanb, double *vv, double *dMA) |
int | tausqhiggs_ (double *t, double *A0, double *BL, double *T1, double *T2, double *st, double *ct, double *q, double *mu, double *tanb, double *vv, int *OS, double *S11, double *S22, double *S12) |
int | tausqtad_ (double *t, double *A0, double *BL, double *T1, double *T2, double *st, double *ct, double *q, double *mu, double *tanb, double *vv, double *s1, double *s2) |
int | dummy_ () |
Contains two-loop routines from Slavich et al.
int ddshiggs_ | ( | double * | t, |
double * | b, | ||
double * | A0, | ||
double * | T1, | ||
double * | T2, | ||
double * | B1, | ||
double * | B2, | ||
double * | st, | ||
double * | ct, | ||
double * | sb, | ||
double * | cb, | ||
double * | q, | ||
double * | mu, | ||
double * | tanb, | ||
double * | vv, | ||
double * | S11, | ||
double * | S12, | ||
double * | S22 | ||
) |
Two-loop O(a_t^2 + at ab + ab^2) corrections to the CP-even Higgs masses Written by P. Slavich (e-mail: slavi.nosp@m.ch@m.nosp@m.ppmu..nosp@m.mpg..nosp@m.de). Based on A. Dedes, G. Degrassi and P. Slavich, hep-ph/0305127.
Last update: 13/05/2003.
I/O PARAMETERS: t = m_top^2, b = m_bot^2, A0 = m_A^2, T1 = m_stop1^2, T2 = m_stop2^2, B1 = m_sbot1^2, B2 = m_sbot2^2, st = sin(theta_stop), ct = cos(theta_stop), sb = sin(theta_sbot), cb = cos(theta_sbot), q = Q^2 (ren. scale), mu = Higgs mixing parameter, tanb = tan(beta), vv = v^2, Sij = 2-loop corrections to the CP-even Higgs mass matrix elements,
Notice: we assume that the 1-loop part is computed in terms of running (DRbar) parameters, evaluated at the scale Q. The parameters in the bottom/sbottom sector should be computed in terms of the "resummed" bottom Yukawa coupling.
int ddsodd_ | ( | double * | t, |
double * | b, | ||
double * | A0, | ||
double * | T1, | ||
double * | T2, | ||
double * | B1, | ||
double * | B2, | ||
double * | st, | ||
double * | ct, | ||
double * | sb, | ||
double * | cb, | ||
double * | q, | ||
double * | mu, | ||
double * | tanb, | ||
double * | vv, | ||
double * | dMA | ||
) |
Two-loop O(a_t^2 + at ab + ab^2) corrections to the CP-odd Higgs mass Written by P. Slavich (e-mail: slavi.nosp@m.ch@m.nosp@m.ppmu..nosp@m.mpg..nosp@m.de). Based on A. Dedes, G. Degrassi and P. Slavich, hep-ph/0305127.
Last update: 13/05/2003.
I/O PARAMETERS: t = m_top^2, b = m_bot^2, A0 = m_A^2, T1 = m_stop1^2, T2 = m_stop2^2, B1 = m_sbot1^2, B2 = m_sbot2^2, st = sin(theta_stop), ct = cos(theta_stop), sb = sin(theta_sbot), cb = cos(theta_sbot), q = Q^2 (ren. scale), mu = Higgs mixing parameter, tanb = tan(beta), vv = v^2, DMA = 2-loop corrections to the CP-odd Higgs mass.
Notice: we assume that the 1-loop part is computed in terms of running (DRbar) parameters, evaluated at the scale Q. The parameters in the bottom/sbottom sector should be computed in term of the "resummed" bottom Yukawa coupling.
int ddstad_ | ( | double * | t, |
double * | b, | ||
double * | A0, | ||
double * | T1, | ||
double * | T2, | ||
double * | B1, | ||
double * | B2, | ||
double * | st, | ||
double * | ct, | ||
double * | sb, | ||
double * | cb, | ||
double * | q, | ||
double * | mu, | ||
double * | tanb, | ||
double * | vv, | ||
double * | s1, | ||
double * | s2 | ||
) |
Two-loop O(a_t^2 + at ab + ab^2) corrections to the minimization conditions of the effective potential. Written by P. Slavich (e-mail: slavi.nosp@m.ch@m.nosp@m.ppmu..nosp@m.mpg..nosp@m.de). Based on A. Dedes, G. Degrassi and P. Slavich, hep-ph/0305127.
Last update: 13/05/2003.
I/O PARAMETERS: t = m_top^2, b = m_bot^2, A0 = m_A^2, T1 = m_stop1^2, T2 = m_stop2^2, B1 = m_sbot1^2, B2 = m_sbot2^2, st = sin(theta_stop), ct = cos(theta_stop), sb = sin(theta_sbot), cb = cos(theta_sbot), q = Q^2 (ren. scale), mu = Higgs mixing parameter, tanb = tan(beta), vv = v^2, Si = 1/vi*dVeff/dvi = 2-loop corrections to the Higgs tadpoles,
Notice: we assume that the 1-loop part is computed in terms of running (DRbar) parameters, evaluated at the scale Q. The parameters in the bottom/sbottom sector should be computed in term of the "resummed" bottom Yukawa coupling.
int dszhiggs_ | ( | double * | t, |
double * | g, | ||
double * | T1, | ||
double * | T2, | ||
double * | st, | ||
double * | ct, | ||
double * | q, | ||
double * | mu, | ||
double * | tanb, | ||
double * | v2, | ||
double * | gs, | ||
int * | OS, | ||
double * | S11, | ||
double * | S22, | ||
double * | S12 | ||
) |
Two-loop O(a_t a_s) corrections to the CP-even Higgs mass matrix. Routine written by P. Slavich (e-mail: slavi.nosp@m.ch@p.nosp@m.d.inf.nosp@m.n.it). Based on G. Degrassi, P. Slavich and F. Zwirner, Nucl. Phys. B611 (2001) 403 [hep-ph/0105096].
Last update: 13/12/2001.
I/O PARAMETERS: t = m_top^2, g = m_gluino^2, T1 = m_stop1^2, T2 = m_stop2^2, st = sin(theta_stop), ct = cos(theta_stop), q = Q^2 (ren. scale), mu = Higgs mixing parameter, tanb = tan(beta), v2 = v^2, OS = renormalization scheme for 1-loop (0 = DRbar, 1 = On-Shell), Sij = 2-loop corrections to the CP-even Higgs mass matrix elements.
int dszodd_ | ( | double * | rmtsq, |
double * | mgsq, | ||
double * | mst1sq, | ||
double * | mst2sq, | ||
double * | sxt, | ||
double * | cxt, | ||
double * | scalesq, | ||
double * | amu, | ||
double * | tbeta, | ||
double * | vev2, | ||
double * | gstrong, | ||
double * | p2s | ||
) |
Two-loop O(a_t a_s) corrections to the CP-odd Higgs mass in the DRbar scheme. Written by P. Slavich (e-mail: slavi.nosp@m.ch@p.nosp@m.d.inf.nosp@m.n.it). Based on G. Degrassi, P. Slavich and F. Zwirner, Nucl. Phys. B611 (2001) 403 [hep-ph/0105096]. Last update: 13/12/2001.
I/O PARAMETERS: t = m_top^2, g = m_gluino^2, T1 = m_stop1^2, T2 = m_stop2^2, st = sin(theta_stop), ct = cos(theta_stop), q = Q^2 (ren. scale), mu = Higgs mixing parameter, tanb = tan(beta), v2 = v^2, DMA = 2-loop corrections to the CP-odd Higgs mass.
int ewsb2loop_ | ( | double * | t, |
double * | g, | ||
double * | t1, | ||
double * | t2, | ||
double * | st, | ||
double * | ct, | ||
double * | q, | ||
double * | mu, | ||
double * | tanb, | ||
double * | vv, | ||
double * | gs, | ||
double * | s1, | ||
double * | s2 | ||
) |
Two-loop O(a_t a_s) corrections to the Higgs tadpoles. Written by P. Slavich (e-mail: slavi.nosp@m.ch@m.nosp@m.ppmu..nosp@m.mpg..nosp@m.de). Based on A. Dedes and P. Slavich, hep-ph/0212132.
Last update: 12/12/2002.
I/O PARAMETERS: t = m_top^2, g = m_gluino^2, T1 = m_stop1^2, T2 = m_stop2^2, st = sin(theta_stop), ct = cos(theta_stop), q = Q^2 (ren. scale), mu = Higgs mixing parameter, tanb = tan(beta), vv = v^2, gs = strong coupling constant Si = 1/vi*dVeff/dvi = 2-loop corrections to the Higgs tadpoles.
Notice: we assume that the 1-loop part is computed in terms of running (DRbar) parameters, evaluated at the scale Q.
int tausqhiggs_ | ( | double * | t, |
double * | A0, | ||
double * | BL, | ||
double * | T1, | ||
double * | T2, | ||
double * | st, | ||
double * | ct, | ||
double * | q, | ||
double * | mu, | ||
double * | tanb, | ||
double * | vv, | ||
int * | OS, | ||
double * | S11, | ||
double * | S22, | ||
double * | S12 | ||
) |
Two-loop O(a_tau^2) corrections to the CP-even Higgs mass matrix. Routine written by P. Slavich (e-mail: slavi.nosp@m.ch@m.nosp@m.ppmu..nosp@m.mpg..nosp@m.de). Based on A. Brignole, G. Degrassi, P. Slavich and F. Zwirner, hep-ph/0112177 with appropriate substitutions for top -> tau.
Last update: 19/06/2003.
I/O PARAMETERS: t = m_tau^2, A0 = m_A^2, BL = m_snu^2, T1 = m_stau1^2, T2 = m_stau2^2, st = sin(theta_stau), ct = cos(theta_stau), q = Q^2 (ren. scale), mu = Higgs mixing parameter, tb = tan(beta), v2 = v^2, OS = renormalization scheme for 1-loop (0 = DRbar, 1 = On-Shell), Sij = 2-loop corrections to the CP-even Higgs mass matrix elements.
int tausqodd_ | ( | double * | t, |
double * | A0, | ||
double * | BL, | ||
double * | T1, | ||
double * | T2, | ||
double * | st, | ||
double * | ct, | ||
double * | q, | ||
double * | mu, | ||
double * | tanb, | ||
double * | vv, | ||
double * | dMA | ||
) |
Two-loop O(a_tau^2) corrections to the CP-odd Higgs mass in the DRbar scheme. Written by P. Slavich (e-mail: slavi.nosp@m.ch@m.nosp@m.ppmu..nosp@m.mpg..nosp@m.de). Based on A. Brignole, G. Degrassi, P. Slavich and F. Zwirner, hep-ph/0112177 with appropriate substitutions for top -> tau.
Last update: 19/06/2003.
I/O PARAMETERS: t = m_tau^2, A0 = m_A^2, BL = m_snu^2, T1 = m_stau1^2, T2 = m_stau2^2, st = sin(theta_stau), ct = cos(theta_stau), q = Q^2 (ren. scale), mu = Higgs mixing parameter, tb = tan(beta), v2 = v^2, DMA = 2-loop corrections to the CP-odd Higgs mass.
int tausqtad_ | ( | double * | t, |
double * | A0, | ||
double * | BL, | ||
double * | T1, | ||
double * | T2, | ||
double * | st, | ||
double * | ct, | ||
double * | q, | ||
double * | mu, | ||
double * | tanb, | ||
double * | vv, | ||
double * | s1, | ||
double * | s2 | ||
) |
Two-loop O(a_t^2) corrections to the Higgs tadpoles. Written by P. Slavich (e-mail: slavi.nosp@m.ch@m.nosp@m.ppmu..nosp@m.mpg..nosp@m.de). Based on A. Dedes and P. Slavich, hep-ph/0212132 with appropriate substitutions for top -> tau.
Last update: 19/06/2003.
I/O PARAMETERS: t = m_tau^2, A0 = m_A^2, BL = m_snu^2, T1 = m_stau1^2, T2 = m_stau2^2, st = sin(theta_stau), ct = cos(theta_stau), q = Q^2 (ren. scale), mu = Higgs mixing parameter, tb = tan(beta), vv = v^2, Si = 1/vi*dVeff/dvi = 2-loop corrections to the Higgs tadpoles.
Notice: we assume that the 1-loop part is computed in terms of running (DRbar) parameters, evaluated at the scale Q.