def inplace_condition_on_cov(self,Sigma_xy,Sigma_yy,my_prediction,obs_distn):
'cov version useful for linearized approximations; compare to inplace_condition_on for the idea'
self.Sigma, self.mu # make sure we are in distribution form
self._mu += Sigma_xy.dot(solve_psd(Sigma_yy+obs_distn.Sigma,obs_distn.mu - my_prediction))
self._Sigma -= Sigma_xy.dot(solve_psd(Sigma_yy+obs_distn.Sigma,Sigma_xy.T))
self._J = self._h = None # invalidate
return self
def compute_Xs(Gamma_i,Gamma_im1,Oi,Oim1):
Xihat = solve_psd(Gamma_i.T.dot(Gamma_i), Gamma_i.T.dot(Oi))
Xip1hat = solve_psd(Gamma_im1.T.dot(Gamma_im1), Gamma_im1.T.dot(Oim1))
return Xihat, Xip1hat
def compute_Xs(Gamma_i,Gamma_im1,Oi,Oim1):
Xihat = solve_psd(Gamma_i.T.dot(Gamma_i), Gamma_i.T.dot(Oi))
Xip1hat = solve_psd(Gamma_im1.T.dot(Gamma_im1), Gamma_im1.T.dot(Oim1))
return Xihat, Xip1hat
def h(self):
if self._h is None:
self._h = solve_psd(self._Sigma,self._mu) if not self._is_diagonal else self._mu / self._Sigma
return self._h
def mu(self):
if self._mu is None:
self._mu = solve_psd(self._J,self._h) if not self._is_diagonal else self._h / self._J
return self._mu
