def partition_functions(psis, z, F, neighbors=None, iat_method=DEFAULT_IAT): """Estimates the asymptotic variance of the partition function (normalization constant) for each window. To get an estimate of the autocovariance of the free energy for each window, multiply the autocovariance of window :math:`i` by :math:` (k_B T / z_i)^2`. Parameters ---------- psis : 3D data structure The values of the bias functions evaluated each window and timepoint. See `datastructures <../datastructures.html#data-from-sampling>`__ for more information. z : 1D array Array containing the normalization constants F : 2D array Overlap matrix for the first EMUS iteration. neighbors : 2D array, optional List showing which windows neighbor which. See neighbors_harmonic in usutils for explanation. iat_method : string, optional Method used to estimate autocorrelation time. See the documentation above. Returns ------- autocovars : ndarray Array of length L (no. windows) where the i'th value corresponds to the autocovariance estimate for :math:`z_i` z_var_contribs : ndarray Two dimensional array, where element i,j corresponds to window j's contribution to the autocovariance of window i. z_var_iats : ndarray Two dimensional array, where element i,j corresponds to the autocorrelation time associated with window j's contribution to the autocovariance of window i. """ iat_routine = ac._get_iat_method(iat_method) L = len(z) z_var_contribs = np.zeros((L, L)) z_var_iats = np.zeros((L, L)) if neighbors is None: # If no neighborlist, assume all windows neighbor neighbors = np.outer(np.ones(L), range(L)).astype(int) groupInv = lm.groupInverse(np.eye(L) - F) # Calculate the partial derivatives of z . # (i,j,k)'th element is partial of z_k w.r.t. F_ij dzdFij = np.outer(z, groupInv).reshape((L, L, L)) # Iterate over windows, getting err contribution from sampling in each for i, psi_i in enumerate(psis): # Data cleaning psi_i_arr = np.array(psi_i) Lneighb = len(neighbors[i]) # Number of neighbors # Normalize psi_j(x_i^t) for all j psi_sum = np.sum(psi_i_arr, axis=1) normedpsis = np.zeros(psi_i_arr.shape) # psi_j / sum_k psi_k for j in xrange(Lneighb): normedpsis[:, j] = psi_i_arr[:, j] / psi_sum # Calculate contribution to as. err. for each z_k for k in xrange(L): dzkdFij = dzdFij[:, :, k] err_t_series = np.dot(normedpsis, dzkdFij[i][neighbors[i]]) iat, mn, sigma = iat_routine(err_t_series) z_var_contribs[k, i] = sigma * sigma z_var_iats[k, i] = iat autocovars = np.sum(z_var_contribs, axis=1) return autocovars, z_var_contribs, z_var_iats
def emus_iter(psis, Avals=None, neighbors=None, return_iats = False,iat_method=DEFAULT_IAT): """Performs one step of the the EMUS iteration. Parameters ---------- psis : 3D data structure The values of the bias functions evaluated each window and timepoint. See `datastructures <../datastructures.html#data-from-sampling>`__ for more information. Avals : 2D array-like, optional Weights in front of :math:`\psi` in the overlap matrix. neighbors : 2D array-like, optional List showing which windows neighbor which. See neighbors_harmonic in usutils. return_iats : bool, optional Whether or not to calculate integrated autocorrelation times of :math:`\psi_ii^*` for each window. iat_method : string, optional Routine to use for calculating said iats. Accepts 'ipce', 'acor', and 'icce'. Returns ------- z : 1D array Normalization constants for each window F : 2D array The overlap matrix constructed for the eigenproblem. iats : 1D array If return_iats chosen, returns the iats that have been estimated. """ # Initialize variables L = len(psis) # Number of windows F = np.zeros((L,L)) # Initialize F Matrix # Take care of defaults.. if return_iats: iats = np.ones(L) iatroutine = ac._get_iat_method(iat_method) if Avals is None: Avals = np.ones((L,L)) if neighbors is None: neighbors = np.outer(np.ones(L),range(L)).astype(int) for i in xrange(L): nbrs_i = neighbors[i] A_nbs = Avals[i][nbrs_i] nbr_index = list(nbrs_i).index(i) Fi_out = calculate_Fi(psis[i],nbr_index,A_nbs,return_iats) if return_iats: Fi, trajs = Fi_out iats[i] = iatroutine(trajs[nbr_index])[0] else: Fi = Fi_out # Unpack the Neighbor list F[i] = unpackNbrs(Fi,nbrs_i,L) z = lm.stationary_distrib(F) if return_iats: return z, F, iats else: return z, F
def _calculate_acovar(psis, dBdF, gdata=None, dBdg=None, neighbors=None, iat_method=DEFAULT_IAT): """ Estimates the autocovariance and autocorrelation times for each window's contribution to the autocovariance of some observable B. Parameters ---------- psis : 3D data structure The values of the bias functions evaluated each window and timepoint. See `datastructures <../datastructures.html#data-from-sampling>`__ for more information. dBdF : array-like Two dimensional array, where element :math:`i,j` is the derivative of the estimate of B with respect to :math:`F_{ij}` gdata : array-like, optional Three dimensional data structure containing data from various observables. The first index n dBdg : array-like, optional Two dimensional array, where element :math:`n,j` is the derivative of the estimate of B with respect to :math:`gn_j^*`. Returns ------- """ L = len(psis) if gdata is not None: if len(gdata) != len(dBdg): raise ValueError( "Function data provided is mismatched with derivatives: respective sizes are ", np.shape(gdata), " and ", np.shape(dBdg), ) if neighbors is None: neighbors = np.outer(np.ones(L), range(L)).astype(int) dBdF = np.array(dBdF) iat_routine = ac._get_iat_method(iat_method) sigmas = np.zeros(L) taus = np.zeros(L) for i, psi_i in enumerate(psis): nbrs_i = neighbors[i] denom_i = 1.0 / np.sum(psi_i, axis=1) errtraj = psi_i * np.transpose([denom_i]) Fi = np.average(errtraj, axis=0) errtraj = np.dot((psi_i * np.transpose([denom_i]) - Fi), dBdF[i, nbrs_i]) if gdata is not None: for n, g_n in enumerate(gdata): g_ni = g_n[i] dBdg_n = dBdg[n] g_ni_wtd = g_ni * denom_i errtraj += dBdg_n[i] * (g_ni_wtd - np.average(g_ni_wtd)) tau, mean, sigma = iat_routine(errtraj) taus[i] = tau sigmas[i] = sigma return taus, sigmas ** 2