def get_second_order_coeff_fi_from_fo(self, psi_x, psi_w, psi_q, d_first, d_second):
gxtp1, gxt, gxtm1, gwtp1, gwpt, gq = d_first
nx = psi_x.shape[1]
nw = psi_w.shape[1]
gxtp1xtp1 = d_second[0]
gxtp1xt = d_second[1]
gxtp1xtm1 = d_second[2]
gxtp1wtp1 = d_second[3]
gxtp1wt = d_second[4]
gxtp1q = d_second[5]
gxtxtp1 = d_second[6]
gxtxt = d_second[7]
gxtxtm1 = d_second[8]
gxtwtp1 = d_second[9]
gxtwt = d_second[10]
gxtq = d_second[11]
gxtm1xtp1 = d_second[12]
gxtm1xt = d_second[13]
gxtm1xtm1 = d_second[14]
gxtm1wtp1 = d_second[15]
gxtm1wt = d_second[16]
gxtm1q = d_second[17]
gwtp1xtp1 = d_second[18]
gwtp1xt = d_second[19]
gwtp1xtm1 = d_second[20]
gwtp1wtp1 = d_second[21]
gwtp1wt = d_second[22]
gwtp1q = d_second[23]
gwtxtp1 = d_second[24]
gwtxt = d_second[25]
gwtxtm1 = d_second[26]
gwtwtp1 = d_second[27]
gwtwt = d_second[28]
gwtq = d_second[29]
gqxtp1 = d_second[30]
gqxt = d_second[31]
gqxtm1 = d_second[32]
gqwtp1 = d_second[33]
gqwt = d_second[34]
gqq = d_second[35]
I_n = np.eye(nx)
I_n_dot_j = [I_n[:, j].reshape((nx, 1)) for j in range(nx)]
psiwkronIndotj_list = [np.kron(psi_w, c) for c in I_n_dot_j]
psiwkronIndotj = np.concatenate(psiwkronIndotj_list, axis=1)
psiwkronIk = np.kron(psi_w, np.eye(nw))
psixkronIn = np.kron(psi_x, np.eye(nx))
psixkronIk = np.kron(psi_x, np.eye(nw))
psiqkronIn = np.kron(psi_q, np.eye(nx))
psiqkronIk = np.kron(psi_q, np.eye(nw))
Inkronpsix = np.kron(np.eye(nx), psi_x)
Inkronpsiw = np.kron(np.eye(nx), psi_w)
Inkronpsiq = np.kron(np.eye(nx), psi_q)
Inkronpsixsquared = np.kron(np.eye(nx), np.dot(psi_x, psi_x))
Ikkronpsix = np.kron(np.eye(nw), psi_x)
psixkronpsix = np.kron(psi_x, psi_x)
psixkronpsiw = np.kron(psi_x, psi_w)
psiwkronpsix = np.kron(psi_w, psi_x)
psixkronpsiq = np.kron(psi_x, psi_q)
psiqkronpsix = np.kron(psi_q, psi_x)
psiwkronpsiq = np.kron(psi_w, psi_q)
psiqkronpsiw = np.kron(psi_q, psi_w)
psiwkronpsiw = np.kron(psi_w, psi_w)
# equation for psi_xx
# A psi_xx + gxtp1 psi_xx B + C
# A = gxt + gxtp1 psi_x
# B = (psi_x kron psi_x)
# C =big constant
# vectorized solution:
# [(I_nn kron A) + (B' kron gxtp1)] vec(psi_xx) = -vec(C)
A_for_psixx = gxt + np.dot(gxtp1, psi_x)
B_for_psixx = np.kron(psi_x, psi_x)
C_for_psixx_1 = gxtm1xtm1 + 2*np.dot(gxtm1xt, Inkronpsix)
C_for_psixx_2 = gxtxt + 2*np.dot(gxtxtp1, Inkronpsix) \
+ np.dot(gxtp1xtp1, psixkronpsix)
C_for_psixx_3 = 2*np.dot(gxtm1xtp1, Inkronpsixsquared)
C_for_psixx = C_for_psixx_1 + np.dot(C_for_psixx_2, psixkronpsix) \
+ C_for_psixx_3
leftmat = np.kron(np.eye(nx*nx), A_for_psixx) + \
np.kron(B_for_psixx.T, gxtp1)
rightmat = - C_for_psixx.T.flatten() # one dimensional array
# two dimensional column array
rightmat2d = - C_for_psixx.T.reshape(-1, 1)
invleftmat = linalg.inv(leftmat)
vec_psixx_sol = np.dot(invleftmat, rightmat)
vec_psixx_sol2d = np.dot(invleftmat, rightmat2d)
vec_psi_x_x_bysolve = linalg.solve(leftmat, rightmat) # twice as fast
psi_x_x = vec_psixx_sol.reshape((nx, nx*nx), order='F')
psi_x_x_2d = vec_psixx_sol2d.reshape((nx, nx*nx), order='F')
# equation for psi_x_w
# 0= 2*Gammax2*psi_xw + 2*gtm1wt + Gammaxw*(psi_x kron Ik) +
# Gammaxtm1xt*(In kron psi_x)
Gamma_xtp1_2 = gxtp1
Gamma_xtp1_xtp1 = gxtp1xtp1
Gamma_xtp1_wtp1 = 2*gxtp1wtp1 + np.dot(Gamma_xtp1_xtp1, psiwkronIndotj)
Gamma_xtp1_wt = 2*gxtp1wt
Gamma_xtp1_q = 2*gxtp1q + np.dot(Gamma_xtp1_xtp1, psiqkronIn)
Gamma_xt_2 = gxt + np.dot(Gamma_xtp1_2, psi_x)
Gamma_xt_xtp1 = 2*gxtxtp1 + np.dot(Gamma_xtp1_xtp1, psixkronIn)
Gamma_xt_xt = gxtxt + \
np.dot(Gamma_xtp1_2, psi_x_x) + np.dot(Gamma_xt_xtp1, Inkronpsix)
Gamma_xt_wt = 2*gxtwt + \
np.dot(Gamma_xtp1_wt, psixkronIk) + \
np.dot(Gamma_xt_xt, psiwkronIndotj)
Gamma_xtm1_xtp1 = 2*gxtm1xtp1
Gamma_xtm1_xt = 2*gxtm1xt + \
np.dot(Gamma_xt_xt, psixkronIn) + \
np.dot(Gamma_xtm1_xtp1, Inkronpsix)
Gamma_wtp1_wt = np.dot(Gamma_xtp1_wt, psiwkronIk)
A_for_psi_x_w = 2*Gamma_xt_2
b_for_psi_x_w = - 2*gxtm1wt + \
np.dot(Gamma_xt_wt, psixkronIk) + np.dot(Gamma_xtm1_xt, Inkronpsiw)
psi_x_w = np.dot(linalg.inv(A_for_psi_x_w), b_for_psi_x_w)
psi_x_w_bysolve = linalg.solve(A_for_psi_x_w, b_for_psi_x_w)
Gamma_xt_wtp1 = 2*gxtwtp1 + 2 * np.dot(Gamma_xtp1_2, psi_x_w) + np.dot(
Gamma_xtp1_wtp1, psixkronIk) + np.dot(Gamma_xt_xtp1, Inkronpsiw)
Gamma_xtm1_wtp1 = 2*gxtm1wtp1 + \
np.dot(Gamma_xt_wtp1, psixkronIk) + \
np.dot(Gamma_xtm1_xtp1, Inkronpsiw)
Gamma_wt_wtp1 = 2*gwtwtp1 + np.dot(Gamma_xt_wtp1, psiwkronIk)
# equation for psi_x_w
# Gammax2*psi_xw + gww + Gammaxw*(psi_w kron Ik)
A_for_psi_w_w = Gamma_xt_2
b_for_psi_w_w = gwtwt + np.dot(Gamma_xt_wt, psiwkronIk)
psi_w_w = np.dot(linalg.inv(A_for_psi_w_w), b_for_psi_w_w)
psi_w_w_bysolve = linalg.solve(A_for_psi_w_w, b_for_psi_w_w)
# Compute Vxx
utility_sym_mat = sympy.Matrix([self.utility])
du_dx = utility_sym_mat.jacobian([self.xvar_t_sym])
du_dx_nopar = [u.subs(self.par_to_values_dict) for u in list(du_dx)]
du_dx_at_ss = [u.subs(self.normal_and_0_to_ss).subs(
self.ss_solutions_dict) for u in du_dx_nopar]
du_dx_at_ss = utils.matrix2numpyfloat(sympy.Matrix(du_dx_at_ss))
beta = self.beta.subs(self.par_to_values_dict)
Ibetaphi = np.eye(nx) - beta*psi_x
Ibetaphi = utils.matrix2numpyfloat(Ibetaphi)
invIbetaphi = linalg.inv(Ibetaphi)
Vx = np.dot(du_dx_at_ss.T, invIbetaphi)
du_dx = utility_sym_mat.jacobian([self.xvar_t_sym])
du_dx_dx = du_dx.jacobian(self.xvar_t_sym)
du_dx_dx_nopar = du_dx_dx.subs(self.par_to_values_dict)
du_dx_dx_at_ss = du_dx_dx_nopar.subs(
self.normal_and_0_to_ss).subs(self.ss_solutions_dict)
du_dx_dx_at_ss = utils.matrix2numpyfloat(du_dx_dx_at_ss)
# now, vectorize and then transpose the hessian. Becomes 1 \times nx^2
# vector
du_dx_dx_at_ss = du_dx_dx_at_ss.T.reshape(-1, 1).T
Vxx_term_1 = du_dx_dx_at_ss + beta * np.dot(Vx, psi_x_x)
inv_of_Vxx_term_2 = np.eye(nx*nx) - beta*psixkronpsix
inv_of_Vxx_term_2 = utils.matrix2numpyfloat(inv_of_Vxx_term_2)
Vxx = np.dot(Vxx_term_1, linalg.inv(inv_of_Vxx_term_2))
# Gx
Gx_term_1 = np.dot(gxtp1, psi_w) + gwtp1
Gx_term_2 = 2*np.dot(Vx, psi_x_w) + np.dot(Vxx, psixkronpsiw)
Gx_term_2 = utils.matrix2numpyfloat(Gx_term_2)
matGx_term_2 = Gx_term_2.reshape(nw, nx, order='F')
Gx_term_3 = np.dot(Vxx, psiwkronpsix)
Gx_term_3 = utils.matrix2numpyfloat(Gx_term_3)
matGx_term_3 = Gx_term_3.reshape(nx, nw, order='F')
theta = self.theta
Gx = np.dot(Gx_term_1, matGx_term_2 + matGx_term_3.T)/theta
# Equation (50) bearing on determintation of psi_x_q
# coe1 psixq + coe2 psixq coe3 + constant terms = 0
# convert it as
# (I kron coe1) vec(psixq) + (coe3' kron coe2) vec(psixq) = -vec(constant terms)
# [(I kron coe1) + (coe3' kron coe2)] vec(psixq) = -vec(constant terms)
# Solve it as Ax = b
coe1 = 2*Gamma_xt_2 # is 7x7
coe2 = 2*Gamma_xtp1_2 # is 7x7
coe3 = psi_x # is 7x7
# now some constant terms:
eq50cons_1 = 2*gxtm1q + \
np.dot(Gamma_xtm1_xtp1, Inkronpsiq) + \
np.dot(Gamma_xtm1_xt, Inkronpsiq)
eq50cons_2_a = 2*gxtq + 2*np.dot(Gamma_xtp1_q, psi_x)
eq50cons_2_b = np.dot(
Gamma_xt_xtp1, Inkronpsiq) + np.dot(Gamma_xt_xt, psiqkronIn)
eq50cons_2 = np.dot(eq50cons_2_a+eq50cons_2_b, psi_x)
E_w_dist = - np.dot(Vx, psi_w).T/theta
eq50cons_3_list = [np.dot(E_w_dist.T, Gamma_xtm1_wtp1[i, :].reshape(
(nw, nx), order='F')) for i in range(nx)]
eq50cons_3 = np.concatenate(eq50cons_3_list)
eq50cons = eq50cons_1 + eq50cons_2 + eq50cons_3
cons_for_psixq = eq50cons + np.dot(Gx, psi_x)
# (coe3' kron coe2) is 49x49, then the I in (I kron coe1) must be 7x7
A_for_vecpsixq = np.kron(np.eye(nx), coe1) + np.kron(coe3.T, coe2)
b_for_vecpsixq = -cons_for_psixq.reshape((-1, 1))
vec_psixq = linalg.solve(A_for_vecpsixq, b_for_vecpsixq)
psi_x_q = vec_psixq.reshape((nx, nx), order='F')
# equation for psi_w_q:
# A psiwq + constants = 0
# A is 2*Gamma_{x2}
# psiwq is a matrix and so it is constants. We could vectorize the equation as
# (I kron A) vec(psiwq) = - vec(constants)
eq51_coe = 2 * Gamma_xt_2
eq51_cons_1 = 2*gwtq + \
np.dot(Gamma_xtp1_wt, psiqkronIk) + np.dot(Gamma_xt_wt, psiqkronIk)
Gamma_xt_q_1of2 = 2*gxtq + 2 * \
np.dot(Gamma_xtp1_2, psi_x_q) + np.dot(Gamma_xtp1_q, psi_x)
Gamma_xt_q_2of2 = np.dot(
Gamma_xt_xtp1, Inkronpsiq)+np.dot(Gamma_xt_xt, psiqkronIn)
Gamma_xt_q = Gamma_xt_q_1of2 + Gamma_xt_q_2of2
eq51_cons_2 = np.dot(Gamma_xt_q, psi_w)
eq51_cons_3a_list = [np.dot(E_w_dist.T, Gamma_wtp1_wt[i, :].reshape(
(nw, nw), order='F')) for i in range(nx)]
eq51_cons_3a = np.concatenate(eq51_cons_3a_list)
eq51_cons_3b_list = [np.dot(E_w_dist.T, Gamma_wt_wtp1[i, :].reshape(
(nw, nw), order='F')) for i in range(nx)]
eq51_cons_3b = np.concatenate(eq51_cons_3b_list)
eq51_cons_3 = eq51_cons_3a + eq51_cons_3b
eq51_cons = eq51_cons_1 + eq51_cons_2 + eq51_cons_3
cons_for_psiwq = eq51_cons + np.dot(Gx, psi_w)
A_for_vecpsiwq = np.kron(np.eye(nw, nw), eq51_coe)
b_for_vecpsiwq = -cons_for_psiwq.reshape((-1, 1))
vec_psiwq = linalg.solve(A_for_vecpsiwq, b_for_vecpsiwq)
psi_w_q = vec_psiwq.reshape((nx, nw), order='F')
# equation for Vxq:
#du_dx = utility_sym_mat.jacobian([self.xvar_t_sym])
du_dx_dq = du_dx.jacobian(sympy.Matrix([self.q]))
du_dx_dq_nopar = du_dx_dq.subs(self.par_to_values_dict)
du_dx_dq_at_ss = du_dx_dq_nopar.subs(
self.normal_and_0_to_ss).subs(self.ss_solutions_dict)
du_dx_dq_at_ss = utils.matrix2numpyfloat(du_dx_dq_at_ss)
# now, vectorize and then transpose the hessian. Becomes 1 \times nx^2
# vector
du_dx_dq_at_ss = du_dx_dq_at_ss.T.reshape(-1, 1).T
uxq = du_dx_dq_at_ss
Vxiq_term1 = beta*np.dot(Vx, psi_x_q)+0.5 * \
np.dot(Vxx, psixkronpsiq+psiqkronpsix)
Vxq_term2a = 2*np.dot(Vx, psi_x_w) + np.dot(Vxx, psixkronpsiw)
Vxq_term2a = Vxq_term2a.reshape((nw, nx), order='F')
Vxq_term2b = np.dot(Vxx, psiwkronpsix)
Vxq_term2b = Vxq_term2b.reshape((nx, nw), order='F')
Vxq_term2c = np.dot(Vx, psi_w)
Vxq_term2 = -beta*0.5 * \
np.dot(Vxq_term2c, (Vxq_term2a + Vxq_term2b.T))/theta
Vxq_rhs = uxq + Vxiq_term1 + Vxq_term2
Vxq_lhs = np.eye(nx) - beta*psi_x
#Vxq = np.dot(Vxq_rhs, scipy.linalg.inv(Vxq_lhs))
Vxq_rhs = utils.matrix2numpyfloat(Vxq_rhs)
Vxq_lhs = utils.matrix2numpyfloat(Vxq_lhs)
Vxq = linalg.solve(Vxq_lhs.T, Vxq_rhs.T).T # solving A.T x.T = b.T
# equation for psi_q_q
Eq52_coe = Gamma_xtp1_2 + Gamma_xt_2
Eq52_cons_1 = gqq + np.dot(Gamma_xtp1_q + Gamma_xt_q, psi_q)
Gamma_wtp1_q_a = 2*gwtp1q + 2 * \
np.dot(Gamma_xtp1_2, psi_w_q) + np.dot(Gamma_xtp1_wtp1, psiqkronIk)
Gamma_wtp1_q_b = np.dot(
Gamma_xtp1_q, psi_w) + np.dot(Gamma_xt_wtp1, psiqkronIk)
Gamma_wtp1_q = Gamma_wtp1_q_a + Gamma_wtp1_q_b
Eq52_cons_2a = np.dot(Gamma_wtp1_q, E_w_dist)
Gamma_wtp1_wtp1 = gwtp1wtp1 + \
np.dot(Gamma_xtp1_2, psi_w_w)+np.dot(Gamma_xtp1_wtp1, psiwkronIk)
E_wkronw_dist = np.eye(
nw) + np.dot(np.dot(Vx, psi_w).T, np.dot(Vx, psi_w))/(theta**2)
E_wkronw_dist = E_wkronw_dist.reshape((-1, 1))
Eq52_cons_2b = np.dot(Gamma_wtp1_wtp1, E_wkronw_dist)
Eq52_cons_2 = Eq52_cons_2a + Eq52_cons_2b
Eq52_cons = Eq52_cons_1 + Eq52_cons_2
cons_psiqq_a = (np.dot(gxtp1, psi_w) + gwtp1)/theta
cons_psiqq_b = 2*np.dot(Vx, psi_w_q) + np.dot(Vxx,
(psiwkronpsiq+psiqkronpsiw)) + 2*np.dot(Vxq, psi_w)
cons_psiqq_ab = np.dot(cons_psiqq_a, cons_psiqq_b.T)
cons_psiqq_c_1 = (
np.dot(Vx, psi_w_w) + np.dot(Vxx, psiwkronpsiw)).reshape((nw, nw), order='F')
cons_psiqq_c = np.dot(cons_psiqq_c_1.T, E_w_dist)
cons_psiqq_ac = np.dot(cons_psiqq_a, cons_psiqq_c)
cons_psiqq_d = np.dot(cons_psiqq_c_1, E_w_dist)
cons_psiqq_ad = np.dot(cons_psiqq_a, cons_psiqq_d)
terms_in_Eq56 = np.dot(
Gx, psi_q) - cons_psiqq_ab - cons_psiqq_ac - cons_psiqq_ad
#terms_in_Eq56 = np.dot(Gx, psi_q) + cons_psiqq_ab - cons_psiqq_ac - cons_psiqq_ad
cons_for_psiqq = Eq52_cons + terms_in_Eq56
cons_for_psiqq = utils.matrix2numpyfloat(cons_for_psiqq)
psi_q_q = linalg.solve(Eq52_coe, -cons_for_psiqq)
self.psi_x_x = psi_x_x
self.psi_x_w = psi_x_w
self.psi_x_q = psi_x_q
self.psi_w_w = psi_w_w
self.psi_w_q = psi_w_q
self.psi_q_q = psi_q_q
return psi_x_x, psi_x_w, psi_x_q, psi_w_w, psi_w_q, psi_q_q
def get_first_order_approx_coeff_fi(self, eqs=[], param_vals_d={}, mod='numpy',
return_evaluated_der=False):
beta = self.beta.subs(self.par_to_values_dict)
theta = self.theta
d_first, d_second = self.get_dgdxw12_evaluated_at_ss(mod='numpy')
gxtp1_ss, gx_ss, gxtm1_ss, gwtp1_ss, gwt_ss, gq_ss = d_first
gxtp1_ss = utils.matrix2numpyfloat(gxtp1_ss)
gx_ss = utils.matrix2numpyfloat(gx_ss)
gxtm1_ss = utils.matrix2numpyfloat(gxtm1_ss)
gwtp1_ss = utils.matrix2numpyfloat(gwtp1_ss)
gwt_ss = utils.matrix2numpyfloat(gwt_ss)
nx = gxtp1_ss.shape[1]
# quad_coeffmat_in_eq_for_P is Uhlig's \Psi matrix, sensible
quad_coeffmat_in_eq_for_P = gxtp1_ss
# lin_coeffmat_in_eq_for_P is Uhlig's (-\Gamma) matrix, sensible
lin_coeffmat_in_eq_for_P = gx_ss
# cons_coeffmat_in_eq_for_P is Uhlig's (-\Theta)\Psi matrix, sensible
cons_coeffmat_in_eq_for_P = gxtm1_ss
nnzero = np.zeros((nx, nx))
# this is Uhlig's Chi matrix:
lin_con_mat = np.bmat(
[[-lin_coeffmat_in_eq_for_P, -cons_coeffmat_in_eq_for_P],
[np.identity(nx), nnzero]])
# this is Uhlig's Delta matrix:
quad_mat = np.bmat(
[[quad_coeffmat_in_eq_for_P, nnzero], [nnzero, np.identity(nx)]])
stable_psi_x = utils.solve_quad_matrix_stable_sol_QZ(lin_con_mat,
quad_mat, nx)
# print 'stable_psi_x', stable_psi_x
psi_x = stable_psi_x
psi_w_inner = np.dot(gxtp1_ss, stable_psi_x) + gx_ss
psi_w = - np.dot(np.linalg.inv(psi_w_inner), gwt_ss)
psi_q_inv_term = linalg.inv(np.dot(gxtp1_ss, psi_x) +
gxtp1_ss + gx_ss)
coef_on_dist_E = np.dot(gxtp1_ss, psi_w) + gwtp1_ss
utility_sym_mat = sympy.Matrix([self.utility])
du_dx = utility_sym_mat.jacobian([self.xvar_t_sym])
du_dx_nopar = [u.subs(self.par_to_values_dict) for u in list(du_dx)]
du_dx_at_ss = [u.subs(self.normal_and_0_to_ss).subs(
self.ss_solutions_dict) for u in du_dx_nopar]
du_dx_at_ss = utils.matrix2numpyfloat(sympy.Matrix(du_dx_at_ss))
nx = len(self.xvar_t_sym)
Ibetaphi = np.eye(nx) - beta*psi_x
Ibetaphi = utils.matrix2numpyfloat(Ibetaphi)
invIbetaphi = linalg.inv(Ibetaphi)
Vx = np.dot(du_dx_at_ss.T, invIbetaphi)
E_w_dist = - np.dot(Vx, psi_w).T/theta
vec_for_diag = np.dot(coef_on_dist_E, E_w_dist)
diag_mat = np.diag(vec_for_diag)
psi_q = - np.dot(psi_q_inv_term, gq_ss + diag_mat)
self.psi_x = psi_x
self.psi_w = psi_w
self.psi_q = psi_q
if return_evaluated_der:
return psi_x, psi_w, psi_q, d_first, d_second
else:
return psi_x, psi_w, psi_q
