Source code for TRXASprefitpack.res.res_decay

'''
res_decay:
submodule for residual function and gradient for fitting time delay scan with the
convolution of sum of exponential decay and instrumental response function

:copyright: 2021-2022 by pistack (Junho Lee).
:license: LGPL3.
'''

from typing import Optional, Sequence, Tuple
import numpy as np
from ..mathfun.irf import calc_eta, deriv_eta
from ..mathfun.irf import calc_fwhm, deriv_fwhm
from ..mathfun.irf import hess_fwhm_eta
from ..mathfun.A_matrix import make_A_matrix_gau, make_A_matrix_cauchy, fact_anal_A
from ..mathfun.exp_conv_irf import deriv_exp_conv_gau, deriv_exp_conv_cauchy
from ..mathfun.exp_conv_irf import hess_exp_conv_gau, hess_exp_conv_cauchy
from ..mathfun.exp_conv_irf import deriv_exp_sum_conv_gau, deriv_exp_sum_conv_cauchy

# residual and gradient function for exponential decay model


[docs] def residual_decay(x0: np.ndarray, base: bool, irf: str, tau_mask: Optional[Sequence[np.ndarray]] = None, t: Optional[Sequence[np.ndarray]] = None, intensity: Optional[Sequence[np.ndarray]] = None, eps: Optional[Sequence[np.ndarray]] = None) -> np.ndarray: ''' residual_decay scipy.optimize.least_squares compatible vector residual function for fitting multiple set of time delay scan with the sum of convolution of exponential decay and instrumental response function Args: x0: initial parameter, if irf == 'g','c': * 1st: fwhm_(G/L) * 2nd to :math:`2+N_{scan}`: time zero of each scan * :math:`2+N_{scan}` to :math:`2+N_{scan}+N_{\\tau}`: time constant of each decay component if irf == 'pv': * 1st and 2nd: fwhm_G, fwhm_L * 3rd to :math:`3+N_{scan}`: time zero of each scan * :math:`3+N_{scan}` to :math:`3+N_{scan}+N_{\\tau}`: time constant of each decay component base: whether or not include baseline (i.e. very long lifetime component) irf: shape of instrumental response function * 'g': normalized gaussian distribution, * 'c': normalized cauchy distribution, * 'pv': pseudo voigt profile :math:`(1-\\eta)g(f) + \\eta c(f)` For pseudo voigt profile, the mixing parameter :math:`\\eta(f_G, f_L)` and uniform fwhm paramter :math:`f(f_G, f_L)` are calculated by `calc_eta` and `calc_fwhm` routine tau_mask (sequence of boolean np.ndarray): whether or not include jth time constant in ith dataset fitting (tau_mask[i][j]) If base is True, size of tau_mask[i] should be `num_tau+1`. t: time points for each data set intensity: sequence of intensity of datasets eps: sequence of estimated error of datasets Returns: Residual vector ''' x0 = np.atleast_1d(x0) if irf in ['g', 'c']: num_irf = 1 fwhm = x0[0] else: num_irf = 2 fwhm = calc_fwhm(x0[0], x0[1]) eta = calc_eta(x0[0], x0[1]) num_t0 = 0 count = 0 for d in intensity: num_t0 = d.shape[1] + num_t0 count = count + d.size chi = np.empty(count) tau = x0[num_irf+num_t0:] if not base: k = 1/tau else: k = np.empty(tau.size+1) k[:-1] = 1/tau k[-1] = 0 end = 0 t0_idx = num_irf dset_idx = 0 for ti, d, e in zip(t, intensity, eps): if tau_mask is None: tm = np.ones_like(k, dtype=bool) else: tm = tau_mask[dset_idx] for j in range(d.shape[1]): t0 = x0[t0_idx] if irf == 'g': A = make_A_matrix_gau(ti-t0, fwhm, k[tm]) elif irf == 'c': A = make_A_matrix_cauchy(ti-t0, fwhm, k[tm]) else: A_gau = make_A_matrix_gau(ti-t0, fwhm, k[tm]) A_cauchy = make_A_matrix_cauchy(ti-t0, fwhm, k[tm]) A = A_gau + eta*(A_cauchy-A_gau) c = fact_anal_A(A, d[:, j], e[:, j]) chi[end:end+d.shape[0]] = ((c@A) - d[:, j])/e[:, j] end = end + d.shape[0] t0_idx = t0_idx + 1 dset_idx = dset_idx + 1 return chi
[docs] def res_grad_decay(x0: np.ndarray, num_comp: int, base: bool, irf: str, tau_mask: Optional[np.ndarray] = None, fix_param_idx: Optional[np.ndarray] = None, t: Optional[Sequence[np.ndarray]] = None, intensity: Optional[Sequence[np.ndarray]] = None, eps: Optional[Sequence[np.ndarray]] = None) -> Tuple[np.ndarray, np.ndarray]: ''' res_grad_decay scipy.optimize.minimize compatible scalar residual and its gradient function for fitting multiple set of time delay scan with the sum of convolution of exponential decay and instrumental response function Args: x0: initial parameter, if irf == 'g','c': * 1st: fwhm_(G/L) * 2nd to :math:`2+N_{scan}`: time zero of each scan * :math:`2+N_{scan}` to :math:`2+N_{scan}+N_{\\tau}`: time constant of each decay component if irf == 'pv': * 1st and 2nd: fwhm_G, fwhm_L * 3rd to :math:`3+N_{scan}`: time zero of each scan * :math:`3+N_{scan}` to :math:`3+N_{scan}+N_{\\tau}`: time constant of each decay component num_comp: number of exponential decay component (except base) base: whether or not include baseline (i.e. very long lifetime component) irf: shape of instrumental response function * 'g': normalized gaussian distribution, * 'c': normalized cauchy distribution, * 'pv': pseudo voigt profile :math:`(1-\\eta)g(f) + \\eta c(f)` For pseudo voigt profile, the mixing parameter :math:`\\eta(f_G, f_L)` and uniform fwhm paramter :math:`f(f_G, f_L)` are calculated by `calc_eta` and `calc_fwhm` routine tau_mask (sequence of boolean np.ndarray): whether or not include jth time constant in ith dataset fitting (tau_mask[i][j]) If base is True, size of tau_mask[i] should be `num_tau+1`. fix_param_idx: index for fixed parameter (masked array for `x0`) tau_mask (sequence of boolean np.ndarray): whether or not include jth time constant in ith dataset fitting (tau_mask[i][j]) If base is True, size of tau_mask[i] should be `num_tau+1`. t: time points for each data set intensity: sequence of intensity of datasets eps: sequence of estimated error of datasets Returns: Tuple of scalar residual function :math:`(\\frac{1}{2}\\sum_i {res}^2_i)` and its gradient ''' x0 = np.atleast_1d(x0) if irf in ['g', 'c']: num_irf = 1 fwhm = x0[0] else: num_irf = 2 eta = calc_eta(x0[0], x0[1]) fwhm = calc_fwhm(x0[0], x0[1]) dfwhm_G, dfwhm_L = deriv_fwhm(x0[0], x0[1]) deta_G, deta_L = deriv_eta(x0[0], x0[1]) num_t0 = 0 count = 0 for d in intensity: num_t0 = num_t0 + d.shape[1] count = count + d.size tau = x0[num_irf+num_t0:] if not base: k = 1/tau else: k = np.empty(tau.size+1) k[:-1] = 1/tau k[-1] = 0 num_param = num_irf+num_t0+num_comp chi = np.empty(count) df = np.empty((count, tau.size+num_irf)) grad = np.empty(num_param) end = 0 t0_idx = num_irf dset_idx = 0 for ti, d, e in zip(t, intensity, eps): step = d.shape[0] if tau_mask is None: tm = np.ones_like(k, dtype=bool) else: tm = tau_mask[dset_idx] for j in range(d.shape[1]): t0 = x0[t0_idx] if irf == 'g': A = make_A_matrix_gau(ti-t0, fwhm, k[tm]) elif irf == 'c': A = make_A_matrix_cauchy(ti-t0, fwhm, k[tm]) else: A_gau = make_A_matrix_gau(ti-t0, fwhm, k[tm]) A_cauchy = make_A_matrix_cauchy(ti-t0, fwhm, k[tm]) diff = A_cauchy-A_gau A = A_gau + eta*diff cm = fact_anal_A(A, d[:, j], e[:, j]) chi[end:end+step] = (cm@A-d[:, j])/e[:, j] c = np.zeros_like(k) c[tm] = cm if irf == 'g': grad_tmp = deriv_exp_sum_conv_gau(ti-t0, fwhm, 1/tau, c, base) elif irf == 'c': grad_tmp = deriv_exp_sum_conv_cauchy( ti-t0, fwhm, 1/tau, c, base) else: grad_gau = deriv_exp_sum_conv_gau(ti-t0, fwhm, 1/tau, c, base) grad_cauchy = deriv_exp_sum_conv_cauchy( ti-t0, fwhm, 1/tau, c, base) grad_tmp = grad_gau + eta*(grad_cauchy-grad_gau) grad_tmp = np.einsum('i,ij->ij', 1/e[:, j], grad_tmp) if irf in ['g', 'c']: df[end:end+step, 0] = grad_tmp[:, 1] else: cdiff = (cm@diff)/e[:, j] df[end:end+step, 0] = dfwhm_G*grad_tmp[:, 1]+deta_G*cdiff df[end:end+step, 1] = dfwhm_L*grad_tmp[:, 1]+deta_L*cdiff grad[t0_idx] = -chi[end:end+step]@grad_tmp[:, 0] df[end:end+step, num_irf:] = np.einsum('j,ij->ij', -1/tau**2, grad_tmp[:, 2:]) end = end + step t0_idx = t0_idx + 1 dset_idx = dset_idx+1 mask = np.ones(num_param, dtype=bool) mask[num_irf:num_irf+num_t0] = False grad[mask] = chi@df if fix_param_idx is not None: grad[fix_param_idx] = 0 return np.sum(chi**2)/2, grad
[docs] def res_hess_decay(x0: np.ndarray, num_comp: int, base: bool, irf: str, tau_mask: Optional[np.ndarray] = None, t: Optional[Sequence[np.ndarray]] = None, intensity: Optional[Sequence[np.ndarray]] = None, eps: Optional[Sequence[np.ndarray]] = None) -> np.ndarray: ''' res_hess_decay Hessian for fitting multiple set of time delay scan with the sum of convolution of exponential decay and instrumental response function Args: x0: initial parameter, if irf == 'g','c': * 1st: fwhm_(G/L) * 2nd to :math:`2+N_{scan}`: time zero of each scan * :math:`2+N_{scan}` to :math:`2+N_{scan}+N_{\\tau}`: time constant of each decay component if irf == 'pv': * 1st and 2nd: fwhm_G, fwhm_L * 3rd to :math:`3+N_{scan}`: time zero of each scan * :math:`3+N_{scan}` to :math:`3+N_{scan}+N_{\\tau}`: time constant of each decay component num_comp: number of exponential decay component (except base) base: whether or not include baseline (i.e. very long lifetime component) irf: shape of instrumental response function * 'g': normalized gaussian distribution, * 'c': normalized cauchy distribution, * 'pv': pseudo voigt profile :math:`(1-\\eta)g(f) + \\eta c(f)` For pseudo voigt profile, the mixing parameter :math:`\\eta(f_G, f_L)` and uniform fwhm paramter :math:`f(f_G, f_L)` are calculated by `calc_eta` and `calc_fwhm` routine tau_mask (sequence of boolean np.ndarray): whether or not include jth time constant in ith dataset fitting (tau_mask[i][j]) If base is True, size of tau_mask[i] should be `num_tau+1`. t: time points for each data set intensity: sequence of intensity of datasets eps: sequence of estimated error of datasets Returns: Hessian of scalar residual function :math:`(\\frac{1}{2}\\sum_i {res}^2_i)` based on the seperation scheme ''' x0 = np.atleast_1d(x0) if irf in ['g', 'c']: num_irf = 1 fwhm = x0[0] else: num_irf = 2 eta = calc_eta(x0[0], x0[1]) fwhm = calc_fwhm(x0[0], x0[1]) dfwhm_G, dfwhm_L = deriv_fwhm(x0[0], x0[1]) deta_G, deta_L = deriv_eta(x0[0], x0[1]) hess_fwhm, hess_eta = hess_fwhm_eta(x0[0], x0[1]) num_t0 = 0 for d in intensity: num_t0 = num_t0 + d.shape[1] tau = x0[num_irf+num_t0:] if not base: k = 1/tau else: k = np.empty(tau.size+1) k[:-1] = 1/tau k[-1] = 0 num_param = num_irf+num_t0+num_comp Hcx = np.zeros((tau.size+1*base, tau.size+num_irf+1)) Hcorr = np.zeros((num_param, num_param)) Hx_1st = np.zeros((num_param, num_param)) Hx_2nd = np.zeros((num_param, num_param)) t0_idx = num_irf dset_idx = 0 for ti, d, e in zip(t, intensity, eps): grad_sum = np.zeros((d.shape[0], 1+num_irf+tau.size)) if tau_mask is None: tm = np.ones_like(k, dtype=bool) else: tm = tau_mask[dset_idx] for j in range(d.shape[1]): t0 = x0[t0_idx] if irf == 'g': A = make_A_matrix_gau(ti-t0, fwhm, k) elif irf == 'c': A = make_A_matrix_cauchy(ti-t0, fwhm, k) else: A_gau = make_A_matrix_gau(ti-t0, fwhm, k) A_cauchy = make_A_matrix_cauchy(ti-t0, fwhm, k) diff = A_cauchy-A_gau A = A_gau + eta*diff cm = fact_anal_A(A[tm, :], d[:, j], e[:, j]) chi = (cm@A[tm, :]-d[:, j])/e[:, j] c = np.zeros_like(k) c[tm] = cm dc = np.einsum('ij,j->ij', A, 1/e[:, j]) Hc = dc @ dc.T grad_sum[:, :] = 0 Hcx[:, :] = 0 if irf in ['g', 'c']: for i in range(tau.size): if irf == 'g': tmp_grad = deriv_exp_conv_gau(ti-t0, fwhm, 1/tau[i]) tmp_hess = c[i]*hess_exp_conv_gau(ti-t0, fwhm, 1/tau[i]) else: tmp_grad = deriv_exp_conv_cauchy(ti-t0, fwhm, 1/tau[i]) tmp_hess = c[i]*hess_exp_conv_cauchy(ti-t0, fwhm, 1/tau[i]) tmp_hess[:, 1] = -tmp_hess[:, 1] # d^2 f / d(-t)d(fwhm) tmp_hess[:, 2] = tmp_hess[:, 2]/tau[i]**2 # d^2 f / d(-t)d(1/k) tmp_hess[:, 4] = -tmp_hess[:, 4]/tau[i]**2 # d^2 f / d(fwhm)d(1/k) tmp_hess[:, 5] = (tmp_hess[:, 5]/tau[i]+2*c[i]*tmp_grad[:, 2])/tau[i]**3 # d^2 f / d(1/k)^2 tmp_grad[:, 0] = -tmp_grad[:, 0] # df/d(-t) tmp_grad[:, 2] = -tmp_grad[:, 2]/tau[i]**2 # d f / d(1/k) tmp_grad = np.einsum('ij,i->ij', tmp_grad, 1/e[:, j]) tmp_hess = np.einsum('ij,i->ij', tmp_hess, 1/e[:, j]) tmp_chi_grad = chi@tmp_grad; tmp_chi_hess = chi@tmp_hess # Hcx Hcx[i, 0] = tmp_chi_grad[1] #fwhm Hcx[i, 1] = tmp_chi_grad[0] #t Hcx[i, i+2] = tmp_chi_grad[2] #tau_i # Hx Hx_2nd[0, 0] = Hx_2nd[0, 0] + tmp_chi_hess[3] #d(fwhm)^2 Hx_2nd[0, t0_idx] = Hx_2nd[0, t0_idx] + tmp_chi_hess[1] # dtd(fwhm) Hx_2nd[0, 1+num_t0+i] = Hx_2nd[0, 1+num_t0+i] + tmp_chi_hess[4] #d(fwhm)d(tau) Hx_2nd[t0_idx, t0_idx] = Hx_2nd[t0_idx, t0_idx] + tmp_chi_hess[0] # dt^2 Hx_2nd[t0_idx, 1+num_t0+i] = tmp_chi_hess[2] # dt dtau_i Hx_2nd[1+num_t0+i, 1+num_t0+i] = \ Hx_2nd[1+num_t0+i, 1+num_t0+i] + tmp_chi_hess[5] # d(tau_i)^2 #Jf grad_sum[:, 0] = grad_sum[:, 0]+c[i]*tmp_grad[:, 1] grad_sum[:, 1] = grad_sum[:, 1]+c[i]*tmp_grad[:, 0] grad_sum[:, 2+i] = c[i]*tmp_grad[:, 2] if base: if irf == 'g': tmp_grad = deriv_exp_conv_gau(ti-t0, fwhm, 0) tmp_hess = c[-1]*hess_exp_conv_gau(ti-t0, fwhm, 0) else: tmp_grad = deriv_exp_conv_cauchy(ti-t0, fwhm, 0) tmp_hess = c[-1]*hess_exp_conv_cauchy(ti-t0, fwhm, 0) tmp_hess[:, 1] = -tmp_hess[:, 1] # d^2 f / d(-t)d(fwhm) tmp_grad[:, 0] = -tmp_grad[:, 0] tmp_grad = np.einsum('ij,i->ij', tmp_grad, 1/e[:, j]) tmp_hess = np.einsum('ij,i->ij', tmp_hess, 1/e[:, j]) tmp_chi_grad = chi@tmp_grad; tmp_chi_hess = chi@tmp_hess # Hcx Hcx[-1, 0] = tmp_chi_grad[1] #fwhm Hcx[-1, 1] = tmp_chi_grad[0] #t # Hx Hx_2nd[0, 0] = Hx_2nd[0, 0] + tmp_chi_hess[3] #d(fwhm)^2 Hx_2nd[0, t0_idx] = Hx_2nd[0, t0_idx] + tmp_chi_hess[1] # dtd(fwhm) Hx_2nd[t0_idx, t0_idx] = Hx_2nd[t0_idx, t0_idx] + tmp_chi_hess[0] # dt^2 #Jf grad_sum[:, 0] = grad_sum[:, 0]+c[-1]*tmp_grad[:, 1] grad_sum[:, 1] = grad_sum[:, 1]+c[-1]*tmp_grad[:, 0] else: for i in range(tau.size): tmp_grad_gau = deriv_exp_conv_gau(ti-t0, fwhm, 1/tau[i]) tmp_hess_gau = c[i]*hess_exp_conv_gau(ti-t0, fwhm, 1/tau[i]) tmp_grad_cauchy = deriv_exp_conv_cauchy(ti-t0, fwhm, 1/tau[i]) tmp_hess_cauchy = c[i]*hess_exp_conv_cauchy(ti-t0, fwhm, 1/tau[i]) tmp_hess_gau[:, 1] = -tmp_hess_gau[:, 1] # d^2 f / d(-t)d(fwhm) tmp_hess_gau[:, 2] = tmp_hess_gau[:, 2]/tau[i]**2 # d^2 f / d(-t)d(1/k) tmp_hess_gau[:, 4] = -tmp_hess_gau[:, 4]/tau[i]**2 # d^2 f / d(fwhm)d(1/k) tmp_hess_gau[:, 5] = \ (tmp_hess_gau[:, 5]/tau[i]+2*c[i]*tmp_grad_gau[:, 2])/tau[i]**3 # d^2 f / d(1/k)^2 tmp_grad_gau[:, 0] = -tmp_grad_gau[:, 0] # df/d(-t) tmp_grad_gau[:, 2] = -tmp_grad_gau[:, 2]/tau[i]**2 # d f / d(1/k) tmp_hess_cauchy[:, 1] = -tmp_hess_cauchy[:, 1] # d^2 f / d(-t)d(fwhm) tmp_hess_cauchy[:, 2] = tmp_hess_cauchy[:, 2]/tau[i]**2 # d^2 f / d(-t)d(1/k) tmp_hess_cauchy[:, 4] = -tmp_hess_cauchy[:, 4]/tau[i]**2 # d^2 f / d(fwhm)d(1/k) tmp_hess_cauchy[:, 5] = \ (tmp_hess_cauchy[:, 5]/tau[i]+2*c[i]*tmp_grad_cauchy[:, 2])/tau[i]**3 # d^2 f / d(1/k)^2 tmp_grad_cauchy[:, 0] = -tmp_grad_cauchy[:, 0] # df/d(-t) tmp_grad_cauchy[:, 2] = -tmp_grad_cauchy[:, 2]/tau[i]**2 # d f / d(1/k) tmp_grad = np.zeros((chi.size, 4)) # fwhm_G fwhm_L t tau tmp_hess = np.zeros((chi.size, 10)) ## Construct tmp_grad and tmp_hess for pseudo voigt # gradient tmp_grad[:, 0] = dfwhm_G*\ (tmp_grad_gau[:, 1]+eta*(tmp_grad_cauchy[:, 1]-tmp_grad_gau[:, 1])) + \ deta_G*diff[i, :] tmp_grad[:, 1] = dfwhm_L*\ (tmp_grad_gau[:, 1]+eta*(tmp_grad_cauchy[:, 1]-tmp_grad_gau[:, 1])) + \ deta_L*diff[i, :] tmp_grad[:, 2] = tmp_grad_gau[:, 0]+\ eta*(tmp_grad_cauchy[:, 0]-tmp_grad_gau[:, 0]) tmp_grad[:, 3] = tmp_grad_gau[:, 2]+\ eta*(tmp_grad_cauchy[:, 2]-tmp_grad_gau[:, 2]) #hessian # fwhm_G, fwhm_G tmp_hess[:, 0] = c[i]*hess_fwhm[0]*\ (tmp_grad_gau[:, 1]+ eta*(tmp_grad_cauchy[:, 1]-tmp_grad_gau[:, 1])) + \ dfwhm_G*(dfwhm_G*(tmp_hess_gau[:, 3]+ eta*(tmp_hess_cauchy[:, 3]-tmp_hess_gau[:, 3]))+ 2*c[i]*deta_G*(tmp_grad_cauchy[:, 1]-tmp_grad_gau[:, 1])) + \ c[i]*hess_eta[0]*diff[i, :] # fwhm_L, fwhm_L tmp_hess[:, 4] = c[i]*hess_fwhm[2]*\ (tmp_grad_gau[:, 1]+ eta*(tmp_grad_cauchy[:, 1]-tmp_grad_gau[:, 1])) + \ dfwhm_L*(dfwhm_L*(tmp_hess_gau[:, 3]+ eta*(tmp_hess_cauchy[:, 3]-tmp_hess_gau[:, 3]))+ 2*c[i]*deta_L*(tmp_grad_cauchy[:, 1]-tmp_grad_gau[:, 1])) + \ c[i]*hess_eta[2]*diff[i, :] # fwhm_G, fwhm_L tmp_hess[:, 1] = c[i]*hess_fwhm[1]*\ (tmp_grad_gau[:, 1]+ eta*(tmp_grad_cauchy[:, 1]-tmp_grad_gau[:, 1])) + \ dfwhm_G*(dfwhm_L*(tmp_hess_gau[:, 3]+ eta*(tmp_hess_cauchy[:, 3]-tmp_hess_gau[:, 3]))+ c[i]*deta_L*(tmp_grad_cauchy[:, 1]-tmp_grad_gau[:, 1])) + \ c[i]*hess_eta[1]*diff[i, :] + \ c[i]*deta_G*dfwhm_L*(tmp_grad_cauchy[:, 1]-tmp_grad_gau[:, 1]) # fwhm_G, other tmp_hess[:, 2] = c[i]*deta_G*(tmp_grad_cauchy[:, 0]-tmp_grad_gau[:, 0])+\ dfwhm_G*(tmp_hess_gau[:, 1]+eta*(tmp_hess_cauchy[:, 1]-tmp_hess_gau[:, 1])) tmp_hess[:, 3] = c[i]*deta_G*(tmp_grad_cauchy[:, 2]-tmp_grad_gau[:, 2])+\ dfwhm_G*(tmp_hess_gau[:, 4]+eta*(tmp_hess_cauchy[:, 4]-tmp_hess_gau[:, 4])) # fwhm_L, other tmp_hess[:, 5] = c[i]*deta_L*(tmp_grad_cauchy[:, 0]-tmp_grad_gau[:, 0])+\ dfwhm_L*(tmp_hess_gau[:, 1]+eta*(tmp_hess_cauchy[:, 1]-tmp_hess_gau[:, 1])) tmp_hess[:, 6] = c[i]*deta_L*(tmp_grad_cauchy[:, 2]-tmp_grad_gau[:, 2])+\ dfwhm_L*(tmp_hess_gau[:, 4]+eta*(tmp_hess_cauchy[:, 4]-tmp_hess_gau[:, 4])) # other, other tmp_hess[:, 7] = tmp_hess_gau[:, 0]+\ eta*(tmp_hess_cauchy[:, 0]-tmp_hess_gau[:, 0]) tmp_hess[:, 8] = tmp_hess_gau[:, 2]+\ eta*(tmp_hess_cauchy[:, 2]-tmp_hess_gau[:, 2]) tmp_hess[:, 9] = tmp_hess_gau[:, 5]+\ eta*(tmp_hess_cauchy[:, 5]-tmp_hess_gau[:, 5]) tmp_grad = np.einsum('ij,i->ij', tmp_grad, 1/e[:, j]) tmp_hess = np.einsum('ij,i->ij', tmp_hess, 1/e[:, j]) tmp_chi_grad = chi@tmp_grad; tmp_chi_hess = chi@tmp_hess # Hcx Hcx[i, :num_irf+1] = tmp_chi_grad[:3] #fwhm_(G,L) t Hcx[i, i+1+num_irf] = tmp_chi_grad[3] #tau_i # Hx Hx_2nd[0, 0] = Hx_2nd[0, 0] + tmp_chi_hess[0] #d(fwhm_G)^2 Hx_2nd[0, 1] = Hx_2nd[0, 1] + tmp_chi_hess[1] # d(fwhm_G)d(fwhm_L) Hx_2nd[0, t0_idx] = Hx_2nd[0, t0_idx] + tmp_chi_hess[2] # dtd(fwhm_G) Hx_2nd[0, num_irf+num_t0+i] = \ Hx_2nd[0, num_irf+num_t0+i] + tmp_chi_hess[3] #d(fwhm_G)d(tau) Hx_2nd[1, 1] = Hx_2nd[1, 1] + tmp_chi_hess[4] #d(fwhm_L)^2 Hx_2nd[1, t0_idx] = Hx_2nd[1, t0_idx] + tmp_chi_hess[5] # dtd(fwhm_L) Hx_2nd[1, num_irf+num_t0+i] = \ Hx_2nd[1, num_irf+num_t0+i] + tmp_chi_hess[6] #d(fwhm_L)d(tau) Hx_2nd[t0_idx, t0_idx] = Hx_2nd[t0_idx, t0_idx] + tmp_chi_hess[7] # dt^2 Hx_2nd[t0_idx, num_irf+num_t0+i] = tmp_chi_hess[8] # dt dtau_i Hx_2nd[num_irf+num_t0+i, num_irf+num_t0+i] = \ Hx_2nd[num_irf+num_t0+i, num_irf+num_t0+i] + tmp_chi_hess[9] # d(tau_i)^2 #Jf grad_sum[:, :num_irf+1] = \ grad_sum[:, :num_irf+1]+c[i]*tmp_grad[:, :num_irf+1] grad_sum[:, 1+i+num_irf] = c[i]*tmp_grad[:, 3] if base: tmp_grad_gau = deriv_exp_conv_gau(ti-t0, fwhm, 0) tmp_hess_gau = c[-1]*hess_exp_conv_gau(ti-t0, fwhm, 0) tmp_grad_cauchy = deriv_exp_conv_cauchy(ti-t0, fwhm, 0) tmp_hess_cauchy = c[-1]*hess_exp_conv_cauchy(ti-t0, fwhm, 0) tmp_hess_gau[:, 1] = -tmp_hess_gau[:, 1] # d^2 f / d(-t)d(fwhm) tmp_grad_gau[:, 0] = -tmp_grad_gau[:, 0] # df/d(-t) tmp_hess_cauchy[:, 1] = -tmp_hess_cauchy[:, 1] # d^2 f / d(-t)d(fwhm) tmp_grad_cauchy[:, 0] = -tmp_grad_cauchy[:, 0] # df/d(-t) tmp_grad = np.zeros((chi.size, 3)) # fwhm_G fwhm_L t tau tmp_hess = np.zeros((chi.size, 6)) ## Construct tmp_grad and tmp_hess for pseudo voigt # gradient tmp_grad[:, 0] = dfwhm_G*\ (tmp_grad_gau[:, 1]+eta*(tmp_grad_cauchy[:, 1]-tmp_grad_gau[:, 1])) + \ deta_G*diff[-1, :] tmp_grad[:, 1] = dfwhm_L*\ (tmp_grad_gau[:, 1]+eta*(tmp_grad_cauchy[:, 1]-tmp_grad_gau[:, 1])) + \ deta_L*diff[-1, :] tmp_grad[:, 2] = tmp_grad_gau[:, 0]+\ eta*(tmp_grad_cauchy[:, 0]-tmp_grad_gau[:, 0]) #hessian # fwhm_G, fwhm_G tmp_hess[:, 0] = c[-1]*hess_fwhm[0]*\ (tmp_grad_gau[:, 1]+ eta*(tmp_grad_cauchy[:, 1]-tmp_grad_gau[:, 1])) + \ dfwhm_G*(dfwhm_G*(tmp_hess_gau[:, 3]+ eta*(tmp_hess_cauchy[:, 3]-tmp_hess_gau[:, 3]))+ 2*c[-1]*deta_G*(tmp_grad_cauchy[:, 1]-tmp_grad_gau[:, 1])) + \ c[-1]*hess_eta[0]*diff[-1, :] # fwhm_L, fwhm_L tmp_hess[:, 3] = c[-1]*hess_fwhm[2]*\ (tmp_grad_gau[:, 1]+ eta*(tmp_grad_cauchy[:, 1]-tmp_grad_gau[:, 1])) + \ dfwhm_L*(dfwhm_L*(tmp_hess_gau[:, 3]+ eta*(tmp_hess_cauchy[:, 3]-tmp_hess_gau[:, 3]))+ 2*c[-1]*deta_L*(tmp_grad_cauchy[:, 1]-tmp_grad_gau[:, 1])) + \ c[-1]*hess_eta[2]*diff[-1, :] # fwhm_G, fwhm_L tmp_hess[:, 1] = c[-1]*hess_fwhm[1]*\ (tmp_grad_gau[:, 1]+ eta*(tmp_grad_cauchy[:, 1]-tmp_grad_gau[:, 1])) + \ dfwhm_G*(dfwhm_L*(tmp_hess_gau[:, 3]+ eta*(tmp_hess_cauchy[:, 3]-tmp_hess_gau[:, 3]))+ c[-1]*deta_L*(tmp_grad_cauchy[:, 1]-tmp_grad_gau[:, 1])) + \ c[-1]*hess_eta[1]*diff[-1, :] + \ c[-1]*deta_G*dfwhm_L*(tmp_grad_cauchy[:, 1]-tmp_grad_gau[:, 1]) # fwhm_G, other tmp_hess[:, 2] = c[-1]*deta_G*(tmp_grad_cauchy[:, 0]-tmp_grad_gau[:, 0])+\ dfwhm_G*(tmp_hess_gau[:, 1]+eta*(tmp_hess_cauchy[:, 1]-tmp_hess_gau[:, 1])) # fwhm_L, other tmp_hess[:, 4] = c[-1]*deta_L*(tmp_grad_cauchy[:, 0]-tmp_grad_gau[:, 0])+\ dfwhm_L*(tmp_hess_gau[:, 1]+eta*(tmp_hess_cauchy[:, 1]-tmp_hess_gau[:, 1])) # other, other tmp_hess[:, 5] = tmp_hess_gau[:, 0]+\ eta*(tmp_hess_cauchy[:, 0]-tmp_hess_gau[:, 0]) tmp_grad = np.einsum('ij,i->ij', tmp_grad, 1/e[:, j]) tmp_hess = np.einsum('ij,i->ij', tmp_hess, 1/e[:, j]) tmp_chi_grad = chi@tmp_grad; tmp_chi_hess = chi@tmp_hess # Hcx Hcx[-1, :num_irf+1] = tmp_chi_grad[:3] #fwhm_(G,L) t # Hx Hx_2nd[0, 0] = Hx_2nd[0, 0] + tmp_chi_hess[0] #d(fwhm_G)^2 Hx_2nd[0, 1] = Hx_2nd[0, 1] + tmp_chi_hess[1] # d(fwhm_G)d(fwhm_L) Hx_2nd[0, t0_idx] = Hx_2nd[0, t0_idx] + tmp_chi_hess[2] # dtd(fwhm_G) Hx_2nd[1, 1] = Hx_2nd[1, 1] + tmp_chi_hess[3] #d(fwhm_L)^2 Hx_2nd[1, t0_idx] = Hx_2nd[1, t0_idx] + tmp_chi_hess[4] # dtd(fwhm_L) Hx_2nd[t0_idx, t0_idx] = Hx_2nd[t0_idx, t0_idx] + tmp_chi_hess[5] # dt^2 #Jf grad_sum[:, :num_irf+1] = \ grad_sum[:, :num_irf+1]+c[-1]*tmp_grad[:, :num_irf+1] # IRF independent part Hx_1st_tmp = grad_sum.T @ grad_sum Hcx = Hcx + dc@grad_sum if base: tm_tau = tm[:-1] else: tm_tau = tm tm_2d = np.einsum('i,j->ij', tm_tau, tm_tau) Hc_mask = Hc[tm, :][:, tm] Hcx_mask = np.zeros((Hc_mask.shape[0], 1+num_irf+np.sum(tm_tau))) Hcx_mask[:, :num_irf+1] = Hcx[tm, :num_irf+1] Hcx_mask[:, num_irf+1:] = Hcx[tm, num_irf+1:][:, tm_tau] b = np.linalg.solve(Hc_mask, Hcx_mask) Hcorr_tmp = b.T @ (Hcx_mask) # fwhm Hx_1st[:num_irf, :num_irf] = \ Hx_1st[:num_irf, :num_irf] + Hx_1st_tmp[:num_irf, :num_irf] Hcorr[:num_irf, :num_irf] = \ Hcorr[:num_irf, :num_irf] + Hcorr_tmp[:num_irf, :num_irf] Hx_1st[:num_irf, t0_idx] = Hx_1st_tmp[:num_irf, num_irf] Hcorr[:num_irf, t0_idx] = Hcorr_tmp[:num_irf, num_irf] Hx_1st[:num_irf, num_irf+num_t0:] = \ Hx_1st[:num_irf, num_irf+num_t0:] + \ Hx_1st_tmp[:num_irf, num_irf+1:] Hcorr[:num_irf, num_irf+num_t0:][:, tm_tau] = \ Hcorr[:num_irf, num_irf+num_t0:][:, tm_tau] + \ Hcorr_tmp[:num_irf, num_irf+1:] # t0 Hx_1st[t0_idx, t0_idx] = Hx_1st_tmp[num_irf, num_irf] Hcorr[t0_idx, t0_idx] = Hcorr_tmp[num_irf, num_irf] Hx_1st[t0_idx, num_irf+num_t0:] = \ Hx_1st_tmp[num_irf, 1+num_irf:] Hcorr[t0_idx, num_irf+num_t0:][tm_tau] = \ Hcorr_tmp[num_irf, 1+num_irf:] # tau Hx_1st[num_irf+num_t0:, num_irf+num_t0:] = \ Hx_1st[num_irf+num_t0:, num_irf+num_t0:] + \ Hx_1st_tmp[num_irf+1:, num_irf+1:] Hcorr[num_irf+num_t0:, num_irf+num_t0:][tm_2d] = \ Hcorr[num_irf+num_t0:, num_irf+num_t0:][tm_2d] + \ Hcorr_tmp[num_irf+1:, num_irf+1:].flatten() t0_idx = t0_idx + 1 dset_idx = dset_idx+1 H = Hx_1st + Hx_2nd - Hcorr for i in range(num_param): for j in range(i+1, num_param): H[j, i] = H[i, j] return H
[docs] def residual_decay_same_t0(x0: np.ndarray, base: bool, irf: str, tau_mask: Optional[Sequence[np.ndarray]] = None, t: Optional[Sequence[np.ndarray]] = None, intensity: Optional[Sequence[np.ndarray]] = None, eps: Optional[Sequence[np.ndarray]] = None) -> np.ndarray: ''' residual_decay_same_t0 scipy.optimize.least_squares compatible vector residual function for fitting multiple set of time delay scan with the sum of convolution of exponential decay and instrumental response function Set Time Zero of every time dset in same dataset same Args: x0: initial parameter, if irf == 'g','c': * 1st: fwhm_(G/L) * 2nd to :math:`2+N_{dset}`: time zero of each data set * :math:`2+N_{dset}` to :math:`2+N_{dset}+N_{\\tau}`: time constant of each decay component if irf == 'pv': * 1st and 2nd: fwhm_G, fwhm_L * 3rd to :math:`3+N_{dset}`: time zero of each data set * :math:`3+N_{dset}` to :math:`3+N_{dset}+N_{\\tau}`: time constant of each decay component base: whether or not include baseline (i.e. very long lifetime component) irf: shape of instrumental response function * 'g': normalized gaussian distribution, * 'c': normalized cauchy distribution, * 'pv': pseudo voigt profile :math:`(1-\\eta)g(f) + \\eta c(f)` For pseudo voigt profile, the mixing parameter :math:`\\eta(f_G, f_L)` and uniform fwhm paramter :math:`f(f_G, f_L)` are calculated by `calc_eta` and `calc_fwhm` routine tau_mask (sequence of boolean np.ndarray): whether or not include jth time constant in ith dataset fitting (tau_mask[i][j]) If base is True, size of tau_mask[i] should be `num_tau+1`. t: time points for each data set intensity: sequence of intensity of datasets eps: sequence of estimated error of datasets Returns: Residual vector ''' x0 = np.atleast_1d(x0) if irf in ['g', 'c']: num_irf = 1 fwhm = x0[0] else: num_irf = 2 fwhm = calc_fwhm(x0[0], x0[1]) eta = calc_eta(x0[0], x0[1]) num_dataset = len(t) count = 0 for i in range(num_dataset): count = count + intensity[i].size chi = np.empty(count) tau = x0[num_irf+num_dataset:] if not base: k = 1/tau else: k = np.empty(tau.size+1) k[:-1] = 1/tau k[-1] = 0 end = 0 t0_idx = num_irf dset_idx = 0 for ti, d, e in zip(t, intensity, eps): t0 = x0[t0_idx] if tau_mask is None: tm = np.ones_like(k, dtype=bool) else: tm = tau_mask[dset_idx] if irf == 'g': A = make_A_matrix_gau(ti-t0, fwhm, k[tm]) elif irf == 'c': A = make_A_matrix_cauchy(ti-t0, fwhm, k[tm]) else: A_gau = make_A_matrix_gau(ti-t0, fwhm, k[tm]) A_cauchy = make_A_matrix_cauchy(ti-t0, fwhm, k[tm]) A = A_gau + eta*(A_cauchy-A_gau) for j in range(d.shape[1]): c = fact_anal_A(A, d[:, j], e[:, j]) chi[end:end+d.shape[0]] = ((c@A) - d[:, j])/e[:, j] end = end + d.shape[0] t0_idx = t0_idx + 1 dset_idx = dset_idx + 1 return chi
[docs] def res_grad_decay_same_t0(x0: np.ndarray, num_comp: int, base: bool, irf: str, tau_mask: Optional[np.ndarray] = None, fix_param_idx: Optional[np.ndarray] = None, t: Optional[Sequence[np.ndarray]] = None, intensity: Optional[Sequence[np.ndarray]] = None, eps: Optional[Sequence[np.ndarray]] = None) -> Tuple[np.ndarray, np.ndarray]: ''' res_grad_decay_same_t0 scipy.optimize.minimize compatible scalar residual and its gradient function for fitting multiple set of time delay scan with the sum of convolution of exponential decay and instrumental response function Args: x0: initial parameter, if irf == 'g','c': * 1st: fwhm_(G/L) * 2nd to :math:`2+N_{dset}`: time zero of each dataset * :math:`2+N_{dset}` to :math:`2+N_{dset}+N_{\\tau}`: time constant of each decay component if irf == 'pv': * 1st and 2nd: fwhm_G, fwhm_L * 3rd to :math:`3+N_{dset}`: time zero of each dataset * :math:`3+N_{dset}` to :math:`3+N_{dset}+N_{\\tau}`: time constant of each decay component num_comp: number of exponential decay component (except base) base: whether or not include baseline (i.e. very long lifetime component) irf: shape of instrumental response function * 'g': normalized gaussian distribution, * 'c': normalized cauchy distribution, * 'pv': pseudo voigt profile :math:`(1-\\eta)g(f) + \\eta c(f)` For pseudo voigt profile, the mixing parameter :math:`\\eta(f_G, f_L)` and uniform fwhm paramter :math:`f(f_G, f_L)` are calculated by `calc_eta` and `calc_fwhm` routine tau_mask (sequence of boolean np.ndarray): whether or not include jth time constant in ith dataset fitting (tau_mask[i][j]) If base is True, size of tau_mask[i] should be `num_tau+1`. fix_param_idx: index for fixed parameter (masked array for `x0`) t: time points for each data set intensity: sequence of intensity of datasets eps: sequence of estimated error of datasets Returns: Tuple of scalar residual function :math:`(\\frac{1}{2}\\sum_i {res}^2_i)` and its gradient ''' x0 = np.atleast_1d(x0) if irf in ['g', 'c']: num_irf = 1 fwhm = x0[0] eta = None else: num_irf = 2 eta = calc_eta(x0[0], x0[1]) fwhm = calc_fwhm(x0[0], x0[1]) deta_G, deta_L = deriv_eta(x0[0], x0[1]) dfwhm_G, dfwhm_L = deriv_fwhm(x0[0], x0[1]) num_dataset = len(t) count = 0 for i in range(num_dataset): count = count + intensity[i].size tau = x0[num_irf+num_dataset:num_irf+num_dataset+num_comp] if base: k = np.empty(tau.size+1) k[:-1] = 1/tau k[-1] = 0 else: k = 1/tau num_param = num_irf+num_dataset+num_comp chi = np.empty(count) df = np.zeros((count, num_irf+num_comp)) grad = np.zeros(num_param) end = 0 t0_idx = num_irf dset_idx = 0 for ti, d, e in zip(t, intensity, eps): # initialize step = d.shape[0] t0 = x0[t0_idx] A = np.empty((num_comp+1*base, step)) A_grad_decay = np.empty((num_comp+1*base, step, 3)) grad_decay = np.empty((step, num_comp+2)) if tau_mask is None: tm = np.ones_like(k, dtype=bool) else: tm = tau_mask[dset_idx] # caching if irf == 'g': A = make_A_matrix_gau(ti-t0, fwhm, k) for i in range(num_comp): A_grad_decay[i, :, :] = deriv_exp_conv_gau(ti-t0, fwhm, k[i]) if base: A_grad_decay[-1, :, :] = deriv_exp_conv_gau(ti-t0, fwhm, 0) elif irf == 'c': A = make_A_matrix_cauchy(ti-t0, fwhm, k) for i in range(num_comp): A_grad_decay[i, :, :] = deriv_exp_conv_cauchy(ti-t0, fwhm, k[i]) if base: A_grad_decay[-1, :, :] = deriv_exp_conv_cauchy(ti-t0, fwhm, 0) else: tmp_gau = make_A_matrix_gau(ti-t0, fwhm, k) tmp_cauchy = make_A_matrix_cauchy(ti-t0, fwhm, k) diff = tmp_cauchy-tmp_gau A = tmp_gau + eta*diff for i in range(num_comp): tmp_grad_gau = deriv_exp_conv_gau(ti-t0, fwhm, k[i]) tmp_grad_cauchy = deriv_exp_conv_cauchy(ti-t0, fwhm, k[i]) A_grad_decay[i, :, :] = tmp_grad_gau + eta*(tmp_grad_cauchy-tmp_grad_gau) if base: tmp_grad_gau = deriv_exp_conv_gau(ti-t0, fwhm, 0) tmp_grad_cauchy = deriv_exp_conv_cauchy(ti-t0, fwhm, 0) A_grad_decay[-1, :, :] = tmp_grad_gau + eta*(tmp_grad_cauchy-tmp_grad_gau) for j in range(d.shape[1]): cm = fact_anal_A(A[tm, :], d[:, j], e[:, j]) chi[end:end+step] = (cm@A[tm, :]-d[:, j])/e[:, j] c = np.zeros_like(k) c[tm] = cm grad_decay[:, :2] = \ np.tensordot(c, A_grad_decay[:, :, :2], axes=1) for i in range(num_comp): grad_decay[:, 2+i] = c[i]*A_grad_decay[i, :, 2] grad_decay = np.einsum('i,ij->ij', 1/e[:, j], grad_decay) if irf in ['g', 'c']: df[end:end+step, 0] = grad_decay[:, 1] else: cdiff = (c@diff)/e[:, j] df[end:end+step, 0] = dfwhm_G*grad_decay[:, 1]+deta_G*cdiff df[end:end+step, 1] = dfwhm_L*grad_decay[:, 1]+deta_L*cdiff grad[t0_idx] = grad[t0_idx] -chi[end:end+step]@grad_decay[:, 0] df[end:end+step, num_irf:num_irf+num_comp] = \ np.einsum('j,ij->ij', -1/tau**2, grad_decay[:, 2:]) end = end + step t0_idx = t0_idx + 1 dset_idx = dset_idx + 1 mask = np.ones(num_param, dtype=bool) mask[num_irf:num_irf+num_dataset] = False grad[mask] = chi@df if fix_param_idx is not None: grad[fix_param_idx] = 0 return np.sum(chi**2)/2, grad
[docs] def res_hess_decay_same_t0(x0: np.ndarray, num_comp: int, base: bool, irf: str, tau_mask: Optional[np.ndarray] = None, t: Optional[Sequence[np.ndarray]] = None, intensity: Optional[Sequence[np.ndarray]] = None, eps: Optional[Sequence[np.ndarray]] = None) -> np.ndarray: ''' res_hess_decay_same_t0 Hessian for fitting multiple set of time delay scan with the sum of convolution of exponential decay and instrumental response function Args: x0: initial parameter, if irf == 'g','c': * 1st: fwhm_(G/L) * 2nd to :math:`2+N_{dset}`: time zero of each dataset * :math:`2+N_{dset}` to :math:`2+N_{dset}+N_{\\tau}`: time constant of each decay component if irf == 'pv': * 1st and 2nd: fwhm_G, fwhm_L * 3rd to :math:`3+N_{dset}`: time zero of each dataset * :math:`3+N_{dset}` to :math:`3+N_{dset}+N_{\\tau}`: time constant of each decay component num_comp: number of exponential decay component (except base) base: whether or not include baseline (i.e. very long lifetime component) irf: shape of instrumental response function * 'g': normalized gaussian distribution, * 'c': normalized cauchy distribution, * 'pv': pseudo voigt profile :math:`(1-\\eta)g(f) + \\eta c(f)` For pseudo voigt profile, the mixing parameter :math:`\\eta(f_G, f_L)` and uniform fwhm paramter :math:`f(f_G, f_L)` are calculated by `calc_eta` and `calc_fwhm` routine fix_param_idx: index for fixed parameter (masked array for `x0`) tau_mask (sequence of boolean np.ndarray): whether or not include jth time constant in ith dataset fitting (tau_mask[i][j]) If base is True, size of tau_mask[i] should be `num_tau+1`. t: time points for each data set intensity: sequence of intensity of datasets eps: sequence of estimated error of datasets Returns: Hessian of scalar residual function :math:`(\\frac{1}{2}\\sum_i {res}^2_i)` based on the seperation scheme ''' x0 = np.atleast_1d(x0) if irf in ['g', 'c']: num_irf = 1 fwhm = x0[0] else: num_irf = 2 eta = calc_eta(x0[0], x0[1]) fwhm = calc_fwhm(x0[0], x0[1]) dfwhm_G, dfwhm_L = deriv_fwhm(x0[0], x0[1]) deta_G, deta_L = deriv_eta(x0[0], x0[1]) hess_fwhm, hess_eta = hess_fwhm_eta(x0[0], x0[1]) num_t0 = len(intensity) tau = x0[num_irf+num_t0:] if not base: k = 1/tau else: k = np.empty(tau.size+1) k[:-1] = 1/tau k[-1] = 0 num_param = num_irf+num_t0+num_comp Hcx = np.zeros((tau.size+1*base, tau.size+num_irf+1)) Hcorr = np.zeros((num_param, num_param)) Hx_1st = np.zeros((num_param, num_param)) Hx_2nd = np.zeros((num_param, num_param)) t0_idx = num_irf dset_idx = 0 for ti, d, e in zip(t, intensity, eps): cache_grad = np.zeros((d.shape[0], num_irf+2, k.size)) cache_hess = np.zeros((d.shape[0], ((num_irf+3)*(num_irf+2))//2, k.size)) grad_sum = np.zeros((d.shape[0], 1+num_irf+tau.size)) if tau_mask is None: tm = np.ones_like(k, dtype=bool) else: tm = tau_mask[dset_idx] t0 = x0[t0_idx] if irf == 'g': A = make_A_matrix_gau(ti-t0, fwhm, k) elif irf == 'c': A = make_A_matrix_cauchy(ti-t0, fwhm, k) else: A_gau = make_A_matrix_gau(ti-t0, fwhm, k) A_cauchy = make_A_matrix_cauchy(ti-t0, fwhm, k) diff = A_cauchy-A_gau A = A_gau + eta*diff # caching if irf in ['g', 'c']: for i in range(tau.size): if irf == 'g': tmp_grad = deriv_exp_conv_gau(ti-t0, fwhm, 1/tau[i]) tmp_hess = hess_exp_conv_gau(ti-t0, fwhm, 1/tau[i]) else: tmp_grad = deriv_exp_conv_cauchy(ti-t0, fwhm, 1/tau[i]) tmp_hess = hess_exp_conv_cauchy(ti-t0, fwhm, 1/tau[i]) tmp_hess[:, 1] = -tmp_hess[:, 1] # d^2 f / d(-t)d(fwhm) tmp_hess[:, 2] = tmp_hess[:, 2]/tau[i]**2 # d^2 f / d(-t)d(1/k) tmp_hess[:, 4] = -tmp_hess[:, 4]/tau[i]**2 # d^2 f / d(fwhm)d(1/k) tmp_hess[:, 5] = (tmp_hess[:, 5]/tau[i]+2*tmp_grad[:, 2])/tau[i]**3 # d^2 f / d(1/k)^2 tmp_grad[:, 0] = -tmp_grad[:, 0] # df/d(-t) tmp_grad[:, 2] = -tmp_grad[:, 2]/tau[i]**2 # d f / d(1/k) cache_grad[:, 0, i] = tmp_grad[:, 1] cache_grad[:, 1, i] = tmp_grad[:, 0] cache_grad[:, 2, i] = tmp_grad[:, 2] cache_hess[:, 0, i] = tmp_hess[:, 3] cache_hess[:, 1, i] = tmp_hess[:, 1] cache_hess[:, 2, i] = tmp_hess[:, 4] cache_hess[:, 3, i] = tmp_hess[:, 0] cache_hess[:, 4, i] = tmp_hess[:, 2] cache_hess[:, 5, i] = tmp_hess[:, 5] if base: if irf == 'g': tmp_grad = deriv_exp_conv_gau(ti-t0, fwhm, 0) tmp_hess = hess_exp_conv_gau(ti-t0, fwhm, 0) else: tmp_grad = deriv_exp_conv_cauchy(ti-t0, fwhm, 0) tmp_hess = hess_exp_conv_cauchy(ti-t0, fwhm, 0) tmp_hess[:, 1] = -tmp_hess[:, 1] # d^2 f / d(-t)d(fwhm) tmp_grad[:, 0] = -tmp_grad[:, 0] cache_grad[:, 0, -1] = tmp_grad[:, 1] cache_grad[:, 1, -1] = tmp_grad[:, 0] cache_hess[:, 0, -1] = tmp_hess[:, 3] cache_hess[:, 1, -1] = tmp_hess[:, 1] cache_hess[:, 2, -1] = tmp_hess[:, 0] else: for i in range(tau.size): tmp_grad_gau = deriv_exp_conv_gau(ti-t0, fwhm, 1/tau[i]) tmp_hess_gau = hess_exp_conv_gau(ti-t0, fwhm, 1/tau[i]) tmp_grad_cauchy = deriv_exp_conv_cauchy(ti-t0, fwhm, 1/tau[i]) tmp_hess_cauchy = hess_exp_conv_cauchy(ti-t0, fwhm, 1/tau[i]) tmp_hess_gau[:, 1] = -tmp_hess_gau[:, 1] # d^2 f / d(-t)d(fwhm) tmp_hess_gau[:, 2] = tmp_hess_gau[:, 2]/tau[i]**2 # d^2 f / d(-t)d(1/k) tmp_hess_gau[:, 4] = -tmp_hess_gau[:, 4]/tau[i]**2 # d^2 f / d(fwhm)d(1/k) tmp_hess_gau[:, 5] = \ (tmp_hess_gau[:, 5]/tau[i]+2*tmp_grad_gau[:, 2])/tau[i]**3 # d^2 f / d(1/k)^2 tmp_grad_gau[:, 0] = -tmp_grad_gau[:, 0] # df/d(-t) tmp_grad_gau[:, 2] = -tmp_grad_gau[:, 2]/tau[i]**2 # d f / d(1/k) tmp_hess_cauchy[:, 1] = -tmp_hess_cauchy[:, 1] # d^2 f / d(-t)d(fwhm) tmp_hess_cauchy[:, 2] = tmp_hess_cauchy[:, 2]/tau[i]**2 # d^2 f / d(-t)d(1/k) tmp_hess_cauchy[:, 4] = -tmp_hess_cauchy[:, 4]/tau[i]**2 # d^2 f / d(fwhm)d(1/k) tmp_hess_cauchy[:, 5] = \ (tmp_hess_cauchy[:, 5]/tau[i]+2*tmp_grad_cauchy[:, 2])/tau[i]**3 # d^2 f / d(1/k)^2 tmp_grad_cauchy[:, 0] = -tmp_grad_cauchy[:, 0] # df/d(-t) tmp_grad_cauchy[:, 2] = -tmp_grad_cauchy[:, 2]/tau[i]**2 # d f / d(1/k) ## Construct tmp_grad and tmp_hess for pseudo voigt # gradient cache_grad[:, 0, i] = dfwhm_G*\ (tmp_grad_gau[:, 1]+eta*(tmp_grad_cauchy[:, 1]-tmp_grad_gau[:, 1])) + \ deta_G*diff[i, :] cache_grad[:, 1, i] = dfwhm_L*\ (tmp_grad_gau[:, 1]+eta*(tmp_grad_cauchy[:, 1]-tmp_grad_gau[:, 1])) + \ deta_L*diff[i, :] cache_grad[:, 2, i] = tmp_grad_gau[:, 0]+\ eta*(tmp_grad_cauchy[:, 0]-tmp_grad_gau[:, 0]) cache_grad[:, 3, i] = tmp_grad_gau[:, 2]+\ eta*(tmp_grad_cauchy[:, 2]-tmp_grad_gau[:, 2]) #hessian # fwhm_G, fwhm_G cache_hess[:, 0, i] = hess_fwhm[0]*\ (tmp_grad_gau[:, 1]+ eta*(tmp_grad_cauchy[:, 1]-tmp_grad_gau[:, 1])) + \ dfwhm_G*(dfwhm_G*(tmp_hess_gau[:, 3]+ eta*(tmp_hess_cauchy[:, 3]-tmp_hess_gau[:, 3]))+ 2*deta_G*(tmp_grad_cauchy[:, 1]-tmp_grad_gau[:, 1])) + \ hess_eta[0]*diff[i, :] # fwhm_L, fwhm_L cache_hess[:, 4, i] = hess_fwhm[2]*\ (tmp_grad_gau[:, 1]+ eta*(tmp_grad_cauchy[:, 1]-tmp_grad_gau[:, 1])) + \ dfwhm_L*(dfwhm_L*(tmp_hess_gau[:, 3]+ eta*(tmp_hess_cauchy[:, 3]-tmp_hess_gau[:, 3]))+ 2*deta_L*(tmp_grad_cauchy[:, 1]-tmp_grad_gau[:, 1])) + \ hess_eta[2]*diff[i, :] # fwhm_G, fwhm_L cache_hess[:, 1, i] = hess_fwhm[1]*\ (tmp_grad_gau[:, 1]+ eta*(tmp_grad_cauchy[:, 1]-tmp_grad_gau[:, 1])) + \ dfwhm_G*(dfwhm_L*(tmp_hess_gau[:, 3]+ eta*(tmp_hess_cauchy[:, 3]-tmp_hess_gau[:, 3]))+ deta_L*(tmp_grad_cauchy[:, 1]-tmp_grad_gau[:, 1])) + \ hess_eta[1]*diff[i, :] + \ deta_G*dfwhm_L*(tmp_grad_cauchy[:, 1]-tmp_grad_gau[:, 1]) # fwhm_G, other cache_hess[:, 2, i] = deta_G*(tmp_grad_cauchy[:, 0]-tmp_grad_gau[:, 0])+\ dfwhm_G*(tmp_hess_gau[:, 1]+eta*(tmp_hess_cauchy[:, 1]-tmp_hess_gau[:, 1])) cache_hess[:, 3, i] = deta_G*(tmp_grad_cauchy[:, 2]-tmp_grad_gau[:, 2])+\ dfwhm_G*(tmp_hess_gau[:, 4]+eta*(tmp_hess_cauchy[:, 4]-tmp_hess_gau[:, 4])) # fwhm_L, other cache_hess[:, 5, i] = deta_L*(tmp_grad_cauchy[:, 0]-tmp_grad_gau[:, 0])+\ dfwhm_L*(tmp_hess_gau[:, 1]+eta*(tmp_hess_cauchy[:, 1]-tmp_hess_gau[:, 1])) cache_hess[:, 6, i] = deta_L*(tmp_grad_cauchy[:, 2]-tmp_grad_gau[:, 2])+\ dfwhm_L*(tmp_hess_gau[:, 4]+eta*(tmp_hess_cauchy[:, 4]-tmp_hess_gau[:, 4])) # other, other cache_hess[:, 7, i] = tmp_hess_gau[:, 0]+\ eta*(tmp_hess_cauchy[:, 0]-tmp_hess_gau[:, 0]) cache_hess[:, 8, i] = tmp_hess_gau[:, 2]+\ eta*(tmp_hess_cauchy[:, 2]-tmp_hess_gau[:, 2]) cache_hess[:, 9, i] = tmp_hess_gau[:, 5]+\ eta*(tmp_hess_cauchy[:, 5]-tmp_hess_gau[:, 5]) if base: tmp_grad_gau = deriv_exp_conv_gau(ti-t0, fwhm, 0) tmp_hess_gau = hess_exp_conv_gau(ti-t0, fwhm, 0) tmp_grad_cauchy = deriv_exp_conv_cauchy(ti-t0, fwhm, 0) tmp_hess_cauchy = hess_exp_conv_cauchy(ti-t0, fwhm, 0) tmp_hess_gau[:, 1] = -tmp_hess_gau[:, 1] # d^2 f / d(-t)d(fwhm) tmp_grad_gau[:, 0] = -tmp_grad_gau[:, 0] # df/d(-t) tmp_hess_cauchy[:, 1] = -tmp_hess_cauchy[:, 1] # d^2 f / d(-t)d(fwhm) tmp_grad_cauchy[:, 0] = -tmp_grad_cauchy[:, 0] # df/d(-t) ## Construct tmp_grad and tmp_hess for pseudo voigt # gradient cache_grad[:, 0, -1] = dfwhm_G*\ (tmp_grad_gau[:, 1]+eta*(tmp_grad_cauchy[:, 1]-tmp_grad_gau[:, 1])) + \ deta_G*diff[-1, :] cache_grad[:, 1, -1] = dfwhm_L*\ (tmp_grad_gau[:, 1]+eta*(tmp_grad_cauchy[:, 1]-tmp_grad_gau[:, 1])) + \ deta_L*diff[-1, :] cache_grad[:, 2, -1] = tmp_grad_gau[:, 0]+\ eta*(tmp_grad_cauchy[:, 0]-tmp_grad_gau[:, 0]) #hessian # fwhm_G, fwhm_G cache_hess[:, 0, -1] = hess_fwhm[0]*\ (tmp_grad_gau[:, 1]+ eta*(tmp_grad_cauchy[:, 1]-tmp_grad_gau[:, 1])) + \ dfwhm_G*(dfwhm_G*(tmp_hess_gau[:, 3]+ eta*(tmp_hess_cauchy[:, 3]-tmp_hess_gau[:, 3]))+ 2*deta_G*(tmp_grad_cauchy[:, 1]-tmp_grad_gau[:, 1])) + \ hess_eta[0]*diff[-1, :] # fwhm_L, fwhm_L cache_hess[:, 3, -1] = hess_fwhm[2]*\ (tmp_grad_gau[:, 1]+ eta*(tmp_grad_cauchy[:, 1]-tmp_grad_gau[:, 1])) + \ dfwhm_L*(dfwhm_L*(tmp_hess_gau[:, 3]+ eta*(tmp_hess_cauchy[:, 3]-tmp_hess_gau[:, 3]))+ 2*deta_L*(tmp_grad_cauchy[:, 1]-tmp_grad_gau[:, 1])) + \ hess_eta[2]*diff[-1, :] # fwhm_G, fwhm_L cache_hess[:, 1, -1] = hess_fwhm[1]*\ (tmp_grad_gau[:, 1]+ eta*(tmp_grad_cauchy[:, 1]-tmp_grad_gau[:, 1])) + \ dfwhm_G*(dfwhm_L*(tmp_hess_gau[:, 3]+ eta*(tmp_hess_cauchy[:, 3]-tmp_hess_gau[:, 3]))+ deta_L*(tmp_grad_cauchy[:, 1]-tmp_grad_gau[:, 1])) + \ hess_eta[1]*diff[-1, :] + \ deta_G*dfwhm_L*(tmp_grad_cauchy[:, 1]-tmp_grad_gau[:, 1]) # fwhm_G, other cache_hess[:, 2, -1] = deta_G*(tmp_grad_cauchy[:, 0]-tmp_grad_gau[:, 0])+\ dfwhm_G*(tmp_hess_gau[:, 1]+eta*(tmp_hess_cauchy[:, 1]-tmp_hess_gau[:, 1])) # fwhm_L, other cache_hess[:, 4, -1] = deta_L*(tmp_grad_cauchy[:, 0]-tmp_grad_gau[:, 0])+\ dfwhm_L*(tmp_hess_gau[:, 1]+eta*(tmp_hess_cauchy[:, 1]-tmp_hess_gau[:, 1])) # other, other cache_hess[:, 5, -1] = tmp_hess_gau[:, 0]+\ eta*(tmp_hess_cauchy[:, 0]-tmp_hess_gau[:, 0]) for j in range(d.shape[1]): cm = fact_anal_A(A[tm, :], d[:, j], e[:, j]) chi = (cm@A[tm, :]-d[:, j])/e[:, j] c = np.zeros_like(k) c[tm] = cm dc = np.einsum('ij,j->ij', A, 1/e[:, j]) Hc = dc @ dc.T grad_sum[:, :] = 0 Hcx[:, :] = 0 if irf in ['g', 'c']: for i in range(tau.size): tmp_grad = cache_grad[:, :, i] tmp_hess = c[i]*cache_hess[:, :, i] tmp_grad = np.einsum('ij,i->ij', tmp_grad, 1/e[:, j]) tmp_hess = np.einsum('ij,i->ij', tmp_hess, 1/e[:, j]) tmp_chi_grad = chi@tmp_grad; tmp_chi_hess = chi@tmp_hess # Hcx Hcx[i, 0] = tmp_chi_grad[0] #fwhm Hcx[i, 1] = tmp_chi_grad[1] #t Hcx[i, i+2] = tmp_chi_grad[2] #tau_i # Hx Hx_2nd[0, 0] = Hx_2nd[0, 0] + tmp_chi_hess[0] #d(fwhm)^2 Hx_2nd[0, t0_idx] = Hx_2nd[0, t0_idx] + tmp_chi_hess[1] # dtd(fwhm) Hx_2nd[0, 1+num_t0+i] = Hx_2nd[0, 1+num_t0+i] + tmp_chi_hess[2] #d(fwhm)d(tau) Hx_2nd[t0_idx, t0_idx] = Hx_2nd[t0_idx, t0_idx] + tmp_chi_hess[3] # dt^2 Hx_2nd[t0_idx, 1+num_t0+i] = Hx_2nd[t0_idx, 1+num_t0+i] + tmp_chi_hess[4] # dt dtau_i Hx_2nd[1+num_t0+i, 1+num_t0+i] = \ Hx_2nd[1+num_t0+i, 1+num_t0+i] + tmp_chi_hess[5] # d(tau_i)^2 #Jf grad_sum[:, 0] = grad_sum[:, 0]+c[i]*tmp_grad[:, 0] grad_sum[:, 1] = grad_sum[:, 1]+c[i]*tmp_grad[:, 1] grad_sum[:, 2+i] = c[i]*tmp_grad[:, 2] if base: tmp_grad = cache_grad[:, :, -1] tmp_hess = c[-1]*cache_hess[:, :, -1] tmp_grad = np.einsum('ij,i->ij', tmp_grad, 1/e[:, j]) tmp_hess = np.einsum('ij,i->ij', tmp_hess, 1/e[:, j]) tmp_chi_grad = chi@tmp_grad; tmp_chi_hess = chi@tmp_hess # Hcx Hcx[-1, 0] = tmp_chi_grad[0] #fwhm Hcx[-1, 1] = tmp_chi_grad[1] #t # Hx Hx_2nd[0, 0] = Hx_2nd[0, 0] + tmp_chi_hess[0] #d(fwhm)^2 Hx_2nd[0, t0_idx] = Hx_2nd[0, t0_idx] + tmp_chi_hess[1] # dtd(fwhm) Hx_2nd[t0_idx, t0_idx] = Hx_2nd[t0_idx, t0_idx] + tmp_chi_hess[2] # dt^2 #Jf grad_sum[:, 0] = grad_sum[:, 0]+c[-1]*tmp_grad[:, 0] grad_sum[:, 1] = grad_sum[:, 1]+c[-1]*tmp_grad[:, 1] else: for i in range(tau.size): tmp_grad = cache_grad[:, :, i] tmp_hess = c[i]*cache_hess[:, :, i] tmp_grad = np.einsum('ij,i->ij', tmp_grad, 1/e[:, j]) tmp_hess = np.einsum('ij,i->ij', tmp_hess, 1/e[:, j]) tmp_chi_grad = chi@tmp_grad; tmp_chi_hess = chi@tmp_hess # Hcx Hcx[i, :num_irf+1] = tmp_chi_grad[:3] #fwhm_(G,L) t Hcx[i, i+1+num_irf] = tmp_chi_grad[3] #tau_i # Hx Hx_2nd[0, 0] = Hx_2nd[0, 0] + tmp_chi_hess[0] #d(fwhm_G)^2 Hx_2nd[0, 1] = Hx_2nd[0, 1] + tmp_chi_hess[1] # d(fwhm_G)d(fwhm_L) Hx_2nd[0, t0_idx] = Hx_2nd[0, t0_idx] + tmp_chi_hess[2] # dtd(fwhm_G) Hx_2nd[0, num_irf+num_t0+i] = \ Hx_2nd[0, num_irf+num_t0+i] + tmp_chi_hess[3] #d(fwhm_G)d(tau) Hx_2nd[1, 1] = Hx_2nd[1, 1] + tmp_chi_hess[4] #d(fwhm_L)^2 Hx_2nd[1, t0_idx] = Hx_2nd[1, t0_idx] + tmp_chi_hess[5] # dtd(fwhm_L) Hx_2nd[1, num_irf+num_t0+i] = \ Hx_2nd[1, num_irf+num_t0+i] + tmp_chi_hess[6] #d(fwhm_L)d(tau) Hx_2nd[t0_idx, t0_idx] = Hx_2nd[t0_idx, t0_idx] + tmp_chi_hess[7] # dt^2 Hx_2nd[t0_idx, num_irf+num_t0+i] = Hx_2nd[t0_idx, num_irf+num_t0+i] + \ tmp_chi_hess[8] # dt dtau_i Hx_2nd[num_irf+num_t0+i, num_irf+num_t0+i] = \ Hx_2nd[num_irf+num_t0+i, num_irf+num_t0+i] + tmp_chi_hess[9] # d(tau_i)^2 #Jf grad_sum[:, :num_irf+1] = \ grad_sum[:, :num_irf+1]+c[i]*tmp_grad[:, :num_irf+1] grad_sum[:, 1+i+num_irf] = c[i]*tmp_grad[:, 3] if base: tmp_grad = cache_grad[:, :, -1] tmp_hess = c[-1]*cache_hess[:, :, -1] tmp_grad = np.einsum('ij,i->ij', tmp_grad, 1/e[:, j]) tmp_hess = np.einsum('ij,i->ij', tmp_hess, 1/e[:, j]) tmp_chi_grad = chi@tmp_grad; tmp_chi_hess = chi@tmp_hess # Hcx Hcx[-1, :num_irf+1] = tmp_chi_grad[:3] #fwhm_(G,L) t # Hx Hx_2nd[0, 0] = Hx_2nd[0, 0] + tmp_chi_hess[0] #d(fwhm_G)^2 Hx_2nd[0, 1] = Hx_2nd[0, 1] + tmp_chi_hess[1] # d(fwhm_G)d(fwhm_L) Hx_2nd[0, t0_idx] = Hx_2nd[0, t0_idx] + tmp_chi_hess[2] # dtd(fwhm_G) Hx_2nd[1, 1] = Hx_2nd[1, 1] + tmp_chi_hess[3] #d(fwhm_L)^2 Hx_2nd[1, t0_idx] = Hx_2nd[1, t0_idx] + tmp_chi_hess[4] # dtd(fwhm_L) Hx_2nd[t0_idx, t0_idx] = Hx_2nd[t0_idx, t0_idx] + tmp_chi_hess[5] # dt^2 #Jf grad_sum[:, :num_irf+1] = \ grad_sum[:, :num_irf+1]+c[-1]*tmp_grad[:, :num_irf+1] # IRF independent part Hx_1st_tmp = grad_sum.T @ grad_sum Hcx = Hcx + dc@grad_sum if base: tm_tau = tm[:-1] else: tm_tau = tm tm_2d = np.einsum('i,j->ij', tm_tau, tm_tau) Hc_mask = Hc[tm, :][:, tm] Hcx_mask = np.zeros((Hc_mask.shape[0], 1+num_irf+np.sum(tm_tau))) Hcx_mask[:, :num_irf+1] = Hcx[tm, :num_irf+1] Hcx_mask[:, num_irf+1:] = Hcx[tm, num_irf+1:][:, tm_tau] b = np.linalg.solve(Hc_mask, Hcx_mask) Hcorr_tmp = b.T @ (Hcx_mask) # fwhm Hx_1st[:num_irf, :num_irf] = \ Hx_1st[:num_irf, :num_irf] + Hx_1st_tmp[:num_irf, :num_irf] Hcorr[:num_irf, :num_irf] = \ Hcorr[:num_irf, :num_irf] + Hcorr_tmp[:num_irf, :num_irf] Hx_1st[:num_irf, t0_idx] = Hx_1st[:num_irf, t0_idx] + \ Hx_1st_tmp[:num_irf, num_irf] Hcorr[:num_irf, t0_idx] = Hcorr[:num_irf, t0_idx] + \ Hcorr_tmp[:num_irf, num_irf] Hx_1st[:num_irf, num_irf+num_t0:] = \ Hx_1st[:num_irf, num_irf+num_t0:] + \ Hx_1st_tmp[:num_irf, num_irf+1:] Hcorr[:num_irf, num_irf+num_t0:][:, tm_tau] = \ Hcorr[:num_irf, num_irf+num_t0:][:, tm_tau] + \ Hcorr_tmp[:num_irf, num_irf+1:] # t0 Hx_1st[t0_idx, t0_idx] = Hx_1st[t0_idx, t0_idx] + \ Hx_1st_tmp[num_irf, num_irf] Hcorr[t0_idx, t0_idx] = Hcorr[t0_idx, t0_idx] + \ Hcorr_tmp[num_irf, num_irf] Hx_1st[t0_idx, num_irf+num_t0:] = \ Hx_1st[t0_idx, num_irf+num_t0:] + \ Hx_1st_tmp[num_irf, 1+num_irf:] Hcorr[t0_idx, num_irf+num_t0:][tm_tau] = \ Hcorr[t0_idx, num_irf+num_t0:][tm_tau] + \ Hcorr_tmp[num_irf, 1+num_irf:] # tau Hx_1st[num_irf+num_t0:, num_irf+num_t0:] = \ Hx_1st[num_irf+num_t0:, num_irf+num_t0:] + \ Hx_1st_tmp[num_irf+1:, num_irf+1:] Hcorr[num_irf+num_t0:, num_irf+num_t0:][tm_2d] = \ Hcorr[num_irf+num_t0:, num_irf+num_t0:][tm_2d] + \ Hcorr_tmp[num_irf+1:, num_irf+1:].flatten() t0_idx = t0_idx + 1 dset_idx = dset_idx+1 H = Hx_1st + Hx_2nd - Hcorr for i in range(num_param): for j in range(i+1, num_param): H[j, i] = H[i, j] return H