def update_resnrm(a, u, y0, resnrm):
au = a.dot(u)
b = u.T.dot(au)
yr = u.T.dot(y0)
z = la.solve_lyapunov(b, yr.dot(yr.T))
resnrm.append(la.norm((au - u.dot(b)).dot(z)))
return
def solve_algric(A=None, R=None, W=None, X0=None,
nwtnstps=20, nwtntol=1e-12,
verbose=False):
""" solve algebraic Riccati Equation via Newton iteration
for dense matrices.
The considered Ric Eqn is `A.TX + XA + XRX = W`
"""
XN = X0
nwtns, nwtnres = 0, 999
while nwtns < nwtnstps and nwtnres > nwtntol:
nwtns += 1
fmat = A + np.dot(R, XN)
cw = W + np.dot(XN, np.dot(R, XN))
XNN = spla.solve_lyapunov(fmat.T, cw)
nwtnres = np.linalg.norm(XN - XNN)
XN = XNN
if verbose:
print 'Newton step {0}: norm of update {1}'.format(nwtns, nwtnres)
if verbose:
ricres = np.dot(A.T, XN) + np.dot(XN, A) + \
np.dot(XN, np.dot(R, XN)) - W
print 'Ric Residual: {0}'.format(np.linalg.norm(ricres))
return XN
def _H2P(A, B, analog):
"""Computes the positive-definite P matrix for determining the H2-norm."""
if analog:
P = solve_lyapunov(A, -np.dot(B, B.T)) # AP + PA^T = -BB^T
else:
# Note: discretization is not performed for the user
P = solve_discrete_lyapunov(A, np.dot(B, B.T)) # APA^T - P = -BB^T
return P
def calc_P(self, A, B, Q):
P = sla.solve_lyapunov(self.A.T, -self.Q)
return P
def __init__(self, system):
self.basis = system.unc.basis
self.args = system.args
self.P = sla.solve_lyapunov(self.args.A.T, self.args.Q_lyap)
# record best fit (matrix distance between model and empirical FC) best_dist = 1e10 # objective FC matrices (empirical) FC0_obj = FC_emp[i_sub,i_cond,0,:,:] FCtau_obj = FC_emp[i_sub,i_cond,i_tau,:,:] stop_opt = False i_opt = 0 while not stop_opt: # calculate Jacobian of dynamical system J = -np.eye(N)/tau_x[i_sub,:].mean() + EC # calculate FC0 and FCtau for model FC0 = spl.solve_lyapunov(J,-Sigma) FCtau = np.dot(FC0,spl.expm(J.T*tau)) # matrices of model error Delta_FC0 = FC0_obj-FC0 Delta_FCtau = FCtau_obj-FCtau # calculate error between model and empirical data for FC0 and FC_tau (matrix distance) dist_FC_tmp = 0.5*(np.sqrt((Delta_FC0**2).sum()/(FC0_obj**2).sum())+np.sqrt((Delta_FCtau**2).sum()/(FCtau_obj**2).sum())) # calculate Pearson correlation between model and empirical data for FC0 and FC_tau Pearson_FC_tmp = 0.5*(stt.pearsonr(FC0.reshape(-1),FC0_obj.reshape(-1))[0]+stt.pearsonr(FCtau.reshape(-1),FCtau_obj.reshape(-1))[0]) # record best model parameters if dist_FC_tmp<best_dist: best_dist = dist_FC_tmp
def SOC(num_neurons, p, R):
C = 1.5
# numerical constant
gamma = 3
n = 10
#learning rate
first_quadrant = np.arange(num_neurons / 2)
second_quadrant = first_quadrant + num_neurons / 2
W = np.zeros((num_neurons, num_neurons))
w0 = R / np.sqrt(p * (1 - p) * (1 + gamma**2) / 2)
for i in np.arange(num_neurons):
for j in np.arange(num_neurons):
if i != j:
if np.random.rand() < p and j <= num_neurons / 2:
W[i, j] = w0 / np.sqrt(num_neurons)
elif np.random.rand() < p and j > num_neurons / 2:
W[i, j] = -gamma * w0 / np.sqrt(num_neurons)
W_adj = np.zeros((num_neurons, num_neurons))
W_adj[W.nonzero()] = 1
A = np.random.rand(num_neurons, num_neurons)
A[A < 0.35] = -1
A[A >= 0.35] = 0
A[A < 0] = 1
A = A - np.diag(np.diag(A))
W_adj = W_adj + A - W_adj * A
alpha = max(np.real(npLA.eig(W)[0]))
s = max(C * alpha, alpha + 0.2)
Q = spLA.solve_lyapunov(np.transpose(W - s * np.eye(num_neurons)),
-2 * np.eye(num_neurons))
i = 0
# plt.figure();
# plt.isinteractive = False;
step = 0
# e = 1 / np. trace(Q);
while alpha >= 0.2:
# solve Lyapunov equation
step = step + 1
if np.mod(step, 500) == 0:
print "number of steps: ", step, "; alpha:", alpha
P = spLA.solve_lyapunov(W - s * np.eye(num_neurons),
-2 * np.eye(num_neurons))
deltaW = n * np.dot(Q, P) / np.trace(np.dot(Q, P))
deltaW[:, first_quadrant] = 0
W = W - deltaW
# clipping positive inhibition neurons to zero
W_i = W[:, second_quadrant]
W_i[W_i > 0] = 0
W = np.hstack((W[:, first_quadrant], W_i))
# # impost relative inhibitory strength constraint
W_EE = W[np.ix_(first_quadrant, first_quadrant)]
W_EI = W[np.ix_(first_quadrant, second_quadrant)]
W_IE = W[np.ix_(second_quadrant, first_quadrant)]
W_II = W[np.ix_(second_quadrant, second_quadrant)]
W_EI = -gamma * W_EI * np.mean(W_EE) / np.mean(W_EI)
W_II = -gamma * W_II * np.mean(W_IE) / np.mean(W_II)
W = np.vstack((np.hstack((W_EE, W_EI)), np.hstack((W_IE, W_II))))
# impose density constraints
W[W_adj == 0] = 0
# i = i + 1;
# if np.mod(i,20) == 0:
# plt.clf();
# plt.axis([-5, 5, -5 ,5])
# plt.plot(npLA.eig(W)[0].real,npLA.eig(W)[0].imag,'ro');
# plt.ion();
# plt.draw()
# plt.pause(0.01);
alpha = max(np.real(npLA.eig(W)[0]))
s = max(C * alpha, alpha + 0.2)
Q = spLA.solve_lyapunov(np.transpose(W - s * np.eye(num_neurons)),
-2 * np.eye(num_neurons))
# e = 1 / np.trace(Q);
a_l = npLA.eig(Q)[1][:, np.argmax(npLA.eig(Q)[0])]
a_r = npLA.eig(Q)[1][:, np.argsort(npLA.eig(Q)[0])[-2]]
return W, a_l, a_r, alpha
control, in_vecs, c_matrix = state_controllability(A, B) # Question: Why does A.transpose give the same eigenvectors as the book # and not plain A?? # Answer: The eig function in the numpy library only calculates # the right hand eigenvectors. The Scipy library provides a eig function # in which you can specify whether you want left or right handed eigenvectors. # To determine controllability you need to calculate the input pole vectors # which is dependant on the left eigenvectors. # Calculate eigen vectors and pole vectors val, vec = LA.eig(A, None, 1, 0, 0, 0) n = lin.matrix_rank(c_matrix) P = LA.solve_lyapunov(A, -B * B.T) # Display results print '\nThe transfer function realization is:' print 'G(s) = ' print np.poly1d(G.num[0], variable='s') print "----------------" print np.poly1d(G.den, variable='s') print '\n1) Eigenvalues are: p1 = ', val[0], 'and p2 = ', val[1] print ' with eigenvectors: q1 = ', vec[:, 0], 'and q2 = ', vec[:, 1] print ' Input pole vectors are: up1 = ', in_vecs[0], 'and up2 = ', in_vecs[1] print '\n2) The controlabillity matrix has rank', n, 'and is given as:' print c_matrix
def l1_norm(sys, rtol=1e-6, max_length=2**18):
"""Returns the L1-norm of a linear system within a relative tolerance.
The L1-norm of a (BIBO stable) linear system is the integral of the
absolute value of its impulse response. For unstable systems this will be
infinite. The L1-norm is important because it bounds the worst-case
output of the system for arbitrary inputs within [-1, 1]. In fact,
this worst-case output is achieved by reversing the input which alternates
between -1 and 1 during the intervals where the impulse response is
negative or positive, respectively (in the limit as T -> infinity).
Algorithm adapted from [1]_ following the methods of [2]_. This works by
iteratively refining lower and upper bounds using progressively longer
simulations and smaller timesteps. The lower bound is given by the
absolute values of the discretized response. The upper bound is given by
refining the time-step intervals where zero-crossings may have occurred.
References:
[1] http://www.mathworks.com/matlabcentral/fileexchange/41587-system-l1-norm/content/l1norm.m # noqa: E501
J.F. Whidborne (April 28, 1995).
[2] Rutland, Neil K., and Paul G. Lane. "Computing the 1-norm of the
impulse response of linear time-invariant systems."
Systems & control letters 26.3 (1995): 211-221.
"""
sys = LinearSystem(sys)
if not sys.analog:
raise ValueError("system (%s) must be analog" % sys)
# Setup state-space system and check stability/conditioning
# we will subtract out D and add it back in at the end
A, B, C, D = sys.ss
alpha = np.max(eig(A)[0].real) # eq (28)
if alpha >= 0:
raise ValueError("system (%s) has unstable eigenvalue: %s" % (
sys, alpha))
# Compute a suitable lower-bound for the L1-norm
# using the steady state response, which is equivalent to the
# L1-norm without an absolute value (i.e. just an integral)
G0 = sys.dcgain - sys.D # -C.inv(A).B
# Compute a suitable upper-bound for the L1-norm
# Note this should be tighter than 2*sum(abs(hankel(sys)))
def _normtail(sig, A, x, C):
# observability gramiam when A perturbed by sig
W = solve_lyapunov(A.T + sig*np.eye(len(A)), -C.T.dot(C))
return np.sqrt(x.dot(W).dot(x.T) / 2 / sig) # eq (39)
xtol = -alpha * 1e-4
_, fopt, _, _ = fminbound(
_normtail, 0, -alpha, (A, B.T, C), xtol=xtol, full_output=True)
# Setup parameters for iterative optimization
L, U = abs(G0), fopt
N = 2**4
T = -1 / alpha
while N <= max_length and .5 * (U - L) / L >= rtol: # eq (25)
# Step 1. Improve the lower bound by simulating more.
dt = T / N
dsys = cont2discrete((A, B, C, 0), dt=dt)
Phi = dsys.A
y = impulse(dsys, dt=None, length=N)
abs_y = abs(y[1:])
L_impulse = np.sum(abs_y)
# bound the missing response from t > T from below
L_tail = abs(G0 - np.sum(y)) # eq (33)
L = max(L, L_impulse + L_tail)
# Step 2. Improve the upper bound using refined interval method.
x = _state_impulse(Phi, x0=B, k=N, delay=0) # eq (38)
abs_e = np.squeeze(abs(C.dot(x.T)))
x = x[:-1] # compensate for computing where thresholds crossed
# find intervals that could have zero-crossings and adjust their
# upper bounds (the lower bound is exact for the other intervals)
CTC = C.T.dot(C)
W = solve_lyapunov(A.T, Phi.T.dot(CTC).dot(Phi) - CTC) # eq (36)
AWA = A.T.dot(W.dot(A))
thresh = np.squeeze( # eq (41)
np.sqrt(dt * np.sum(x.dot(AWA) * x, axis=1)))
cross = np.maximum(abs_e[:-1], abs_e[1:]) <= thresh # eq (20)
abs_y[cross] = np.sqrt( # eq (22, 37)
dt * np.sum(x[cross].dot(W) * x[cross], axis=1))
# bound the missing response from t > T from above
_, U_tail, _, _ = fminbound(
_normtail, 0, -alpha, (A, x[-1], C), xtol=xtol, full_output=True)
U_impulse = np.sum(abs_y)
U = max(min(U, U_impulse + U_tail), L)
N *= 2
if U_impulse - L_impulse < U_tail - L_tail: # eq (26)
T *= 2
return (U + L) / 2 + D, .5 * (U - L) / L
def fit(self, X, y=None, SC_mask=None, method='lyapunov', norm_fc=None,
epsilon_EC=0.0005, epsilon_Sigma=0.05, min_val_EC=0.0, max_val_EC=0.4,
n_opt=10000, regul_EC=0, regul_Sigma=0,
true_model=None, verbose=0):
"""
Estimation of MOU parameters (connectivity C, noise covariance Sigma,
and time constant tau_x) with Lyapunov optimization as in: Gilson et al.
Plos Computational Biology (2016).
PARAMETERS:
X: the timeseries data of the system to estimate, shape: T time points x P variables.
y: needed to comply with BaseEstimator (not used here)
SC_mask: mask of known non-zero values for connectivity matrix, for example
estimated by DTI
method: 'lyapunov' or 'moments'
norm_fc: normalization factor for FC. Normalization is needed to avoid high connectivity value that make
the network activity explode. FC is normalized as FC *= 0.5/norm_fc. norm_fc can be specified to be for example
the average over all entries of FC for all subjects or sessions in a given group. If not specified the normalization factor is
the mean of 0-lag covariance matrix of the provided time series ts_emp.
epsilon_EC : learning rate for connectivity (this should be about n_nodes times smaller than epsilon_Sigma).
epsilon_Sigma : learning rate for Sigma (this should be about n_nodes times larger than epsilon_EC).
min_val_EC : minimum value to bound connectivity estimate. This should be zero or slightly negative (too negative limit can bring to an
inhibition dominated system). If the empirical covariance has many negative entries then a slightly negative limit can improve the estimation
accuracy.
max_val_EC : maximum value to bound connectivity estimate. This is useful to avoid large weight that make the system unstable.
If the estimated connectivity saturates toward this value (it usually doesn't happen) it can be increased.
n_opt : number of maximum optimization steps. If final number of iterations reaches this maximum it means the algorithm has not converged.
regul_EC : regularization parameter for connectivity (try a value of 0.5)
regul_Sigma : regularization parameter for Sigma (try a value of 0.001)
true_model: a tuple (true_C, true_S) of true connectivity matrix and noise covariance (for calculating the error of estimation iteratively)
verbose: verbosity level; 0: no output; 1: prints regularization parameters,
estimated \tau_x, used lag \tau for lagged-covariance and evolution of the iterative
optimization with final values of data covariance (Functional Connectivity) fitting;
2: create a diagnostic graphics with cost function over iteration
(distance between true and estimated connectivity is also shown
if true_C is not None) and fitting of data covariance matrix (fitting of connectivity
and Sigma are also shown if true_C and true_S are not None).
RETURN:
C: estimated connectivity matrix, shape [P, P] with null-diagonal
Sigma: estimated noise covariance, shape [P, P]
tau_x: estimated time constant (scalar)
d_fit: a dictionary with diagnostics of the fit; keys are: iterations, distance and correlation
"""
# TODO: raise a warning if the algorithm does not converge
# TODO: look into regularization
# TODO: get rid of comparison with ground truth
# TODO: move SC_make to __init__()
# TODO: make better graphics (deal with axes separation, etc.)
# FIXME: tau_x in Matt origina script is calculated as the mean tau_x over sessions for each subject: why? Is this import?
# TODO: check consistent N between object init and time series X passed to fit()
if not(true_model is None): # if true model is known
true_C = true_model[0] # used later to calculate error iteratively
true_S = true_model[1]
n_T = X.shape[0] # number of time samples
N = X.shape[1] # number of ROIs
d_fit = dict() # a dictionary to store the diagnostics of fit
# mask for existing connections for EC and Sigma
mask_diag = np.eye(N,dtype=bool)
if SC_mask is None:
mask_EC = np.logical_not(mask_diag) # all possible connections except self
else:
mask_EC = SC_mask
mask_Sigma = np.eye(N,dtype=bool) # independent noise
#mask_Sigma = np.ones([N,N],dtype=bool) # coloured noise
if method=='lyapunov':
if verbose>0:
print('regularization:', regul_EC, ';', regul_Sigma)
n_tau = 3 # number of time shifts for FC_emp
v_tau = np.arange(n_tau)
i_tau_opt = 1 # time shift for optimization
min_val_Sigma_diag = 0. # minimal value for Sigma
# FC matrix
ts_emp = X - np.outer(np.ones(n_T), X.mean(0))
FC_emp = np.zeros([n_tau,N,N])
for i_tau in range(n_tau):
FC_emp[i_tau,:,:] = np.tensordot(ts_emp[0:n_T-n_tau,:],ts_emp[i_tau:n_T-n_tau+i_tau,:],axes=(0,0)) / float(n_T-n_tau-1)
# normalize covariances (to ensure the system does not explode)
if norm_fc is None:
norm_fc = FC_emp[0,:,:].mean()
FC_emp *= 0.5/norm_fc
if verbose>0:
print('max FC value (most of the distribution should be between 0 and 1):', FC_emp.mean())
# autocovariance time constant
log_ac = np.log(np.maximum(FC_emp.diagonal(axis1=1,axis2=2),1e-10))
lin_reg = np.polyfit(np.repeat(v_tau,N),log_ac.reshape(-1),1)
tau_x = -1./lin_reg[0]
if verbose>0:
print('inverse of negative slope (time constant):', tau_x)
# optimization
if verbose>0:
print('*opt*')
print('i tau opt:', i_tau_opt)
tau = v_tau[i_tau_opt]
# objective FC matrices (empirical)
FC0_obj = FC_emp[0,:,:]
FCtau_obj = FC_emp[i_tau_opt,:,:]
coef_0 = np.sqrt(np.sum(FCtau_obj**2)) / (np.sqrt(np.sum(FC0_obj**2))+np.sqrt(np.sum(FCtau_obj**2)))
coef_tau = 1. - coef_0
# initial network parameters
EC = np.zeros([N,N])
Sigma = np.eye(N) # initial noise
# best distance between model and empirical data
best_dist = 1e10
best_Pearson = 0.
# record model parameters and outputs
dist_FC_hist = np.zeros([n_opt])*np.nan # FC error = matrix distance
Pearson_FC_hist = np.zeros([n_opt])*np.nan # Pearson corr model/objective
dist_EC_hist = np.zeros([n_opt])*np.nan # FC error = matrix distance
Pearson_EC_hist = np.zeros([n_opt])*np.nan # Pearson corr model/objective
stop_opt = False
i_opt = 0
while not stop_opt:
# calculate Jacobian of dynamical system
J = -np.eye(N)/tau_x + EC
# calculate FC0 and FCtau for model
FC0 = spl.solve_lyapunov(J,-Sigma)
FCtau = np.dot(FC0,spl.expm(J.T*tau))
# calculate error between model and empirical data for FC0 and FC_tau (matrix distance)
err_FC0 = np.sqrt(np.sum((FC0-FC0_obj)**2))/np.sqrt(np.sum(FC0_obj**2))
err_FCtau = np.sqrt(np.sum((FCtau-FCtau_obj)**2))/np.sqrt(np.sum(FCtau_obj**2))
dist_FC_hist[i_opt] = 0.5*(err_FC0+err_FCtau)
if not(true_model is None):
dist_EC_hist[i_opt] = np.sqrt(np.sum((EC-true_C)**2))/np.sqrt(np.sum(true_C**2))
# calculate Pearson corr between model and empirical data for FC0 and FC_tau
Pearson_FC_hist[i_opt] = 0.5*(stt.pearsonr(FC0.reshape(-1),FC0_obj.reshape(-1))[0]+stt.pearsonr(FCtau.reshape(-1),FCtau_obj.reshape(-1))[0])
if not(true_model is None):
Pearson_EC_hist[i_opt] = stt.pearsonr(EC.reshape(-1), true_C.reshape(-1))[0]
# best fit given by best Pearson correlation coefficient for both FC0 and FCtau (better than matrix distance)
if dist_FC_hist[i_opt]<best_dist:
best_dist = dist_FC_hist[i_opt]
best_Pearson = Pearson_FC_hist[i_opt]
i_best = i_opt
EC_best = np.array(EC)
Sigma_best = np.array(Sigma)
FC0_best = np.array(FC0)
FCtau_best = np.array(FCtau)
else:
stop_opt = i_opt>100
# Jacobian update with weighted FC updates depending on respective error
Delta_FC0 = (FC0_obj-FC0)*coef_0
Delta_FCtau = (FCtau_obj-FCtau)*coef_tau
Delta_J = np.dot(np.linalg.pinv(FC0),Delta_FC0+np.dot(Delta_FCtau,spl.expm(-J.T*tau))).T/tau
# update conectivity and noise
EC[mask_EC] += epsilon_EC * (Delta_J - regul_EC*EC)[mask_EC]
EC[mask_EC] = np.clip(EC[mask_EC],min_val_EC,max_val_EC)
Sigma[mask_Sigma] += epsilon_Sigma * (-np.dot(J,Delta_FC0)-np.dot(Delta_FC0,J.T) - regul_Sigma)[mask_Sigma]
Sigma[mask_diag] = np.maximum(Sigma[mask_diag],min_val_Sigma_diag)
# check if end optimization: if FC error becomes too large
if stop_opt or i_opt==n_opt-1:
stop_opt = True
d_fit['iterations'] = i_opt
d_fit['distance'] = best_dist
d_fit['correlation'] = best_Pearson
if verbose>0:
print('stop at step', i_opt, 'with best dist', best_dist, ';best FC Pearson:', best_Pearson)
else:
if (i_opt)%20==0 and verbose>0:
print('opt step:', i_opt)
print('current dist FC:', dist_FC_hist[i_opt], '; current Pearson FC:', Pearson_FC_hist[i_opt])
i_opt += 1
if verbose>1:
# plots
mask_nodiag = np.logical_not(np.eye(N,dtype=bool))
mask_nodiag_and_not_EC = np.logical_and(mask_nodiag,np.logical_not(mask_EC))
mask_nodiag_and_EC = np.logical_and(mask_nodiag,mask_EC)
pp.figure()
gs = GridSpec(2, 3)
if not(true_model is None):
pp.subplot(gs[0,2])
pp.scatter(true_C, EC_best, marker='x')
pp.xlabel('original EC')
pp.ylabel('estimated EC')
pp.text(pp.xlim()[0]+.05, pp.ylim()[1]-.05,
r'$\rho$: ' + str(stt.pearsonr(true_C[mask_EC], EC_best[mask_EC])[0]))
pp.subplot(gs[1,2])
pp.scatter(true_S, Sigma_best,marker='x')
pp.xlabel('original Sigma')
pp.ylabel('estimated Sigma')
pp.text(pp.xlim()[0]+.05, pp.ylim()[1]-.05,
r'$\rho_{diag}$: ' + str(stt.pearsonr(true_S.diagonal(), Sigma_best.diagonal())[0])
+ r'$\rho_{off-diag}$: ' + str(stt.pearsonr(true_S[mask_nodiag], Sigma_best[mask_nodiag])[0])
)
pp.subplot(gs[0,0:2])
pp.plot(range(n_opt),dist_FC_hist, label='distance FC')
pp.plot(range(n_opt),Pearson_FC_hist, label=r'$\rho$ FC')
if not(true_model is None):
pp.plot(range(n_opt),dist_EC_hist, label='distance EC')
pp.plot(range(n_opt),Pearson_EC_hist, label=r'$\rho$ EC')
pp.legend()
pp.xlabel('optimization step')
pp.ylabel('FC error')
pp.subplot(gs[1,0])
pp.scatter(FC0_obj[mask_nodiag_and_not_EC], FC0_best[mask_nodiag_and_not_EC], marker='x', color='k', label='not(SC)')
pp.scatter(FC0_obj[mask_nodiag_and_EC], FC0_best[mask_nodiag_and_EC], marker='.', color='b', label='SC')
pp.scatter(FC0_obj.diagonal(), FC0_best.diagonal(), marker= '.', color='c', label='diagonal')
pp.legend()
pp.xlabel('FC0 emp')
pp.ylabel('FC0 model')
pp.subplot(gs[1,1])
pp.scatter(FCtau_obj[mask_nodiag_and_not_EC], FCtau_best[mask_nodiag_and_not_EC], marker='x', color='k', label='not(SC)')
pp.scatter(FCtau_obj[mask_nodiag_and_EC], FCtau_best[mask_nodiag_and_EC], marker='.', color='b', label='SC')
pp.scatter(FCtau_obj.diagonal(), FCtau_best.diagonal(), marker= '.', color='c', label='diagonal')
pp.xlabel('FCtau emp')
pp.ylabel('FCtau model')
elif method=='moments':
n_tau = 2
ts_emp = X - np.outer(np.ones(n_T), X.mean(0)) # subtract mean
# empirical covariance (0 and 1 lagged)
Q_emp = np.zeros([n_tau, self.n_nodes, self.n_nodes])
for i_tau in range(n_tau):
Q_emp[i_tau, :, :] = np.tensordot(ts_emp[0:n_T-n_tau,:],ts_emp[i_tau:n_T-n_tau+i_tau,:],axes=(0,0)) / float(n_T-n_tau-1)
# Jacobian estimate
J = spl.logm(np.dot(np.linalg.inv(Q_emp[0, :, :]), Q_emp[1, :, :])).T # WARNING: tau is 1 here (if a different one is used C gets divided by tau)
if np.any(np.iscomplex(J)):
J = np.real(J)
print("Warning: complex values in J; casting to real!")
# Sigma estimate
Sigma_best = -np.dot(J, Q_emp[0, :, :])-np.dot(Q_emp[0, :, :], J.T)
# theoretical covariance
Q0 = spl.solve_lyapunov(J, -Sigma_best)
# theoretical 1-lagged covariance
Qtau = np.dot(Q0, spl.expm(J.T)) # WARNING: tau is 1 here (if a different one is used J.T gets multiplied by tau)
# average correlation between empirical and theoretical
d_fit['correlation'] = 0.5 * (stt.pearsonr(Q0.flatten(), Q_emp[0, :, :].flatten())[0] +
stt.pearsonr(Qtau.flatten(), Q_emp[1, :, :].flatten())[0])
tau_x = -J.diagonal().copy()
EC_best = np.zeros([N, N])
EC_best[mask_EC] = J[mask_EC]
elif method=='bayes':
X = X.T # here shape is [ROIs, timepoints] to be consistent with Singh paper
N = X.shape[1] # number of time steps
X -= np.outer(X.mean(axis=1), np.ones(N)) # center the time series
T1 = [np.dot(X[:, i+1:i+2], X[:, i+1:i+2].T) for i in range(N-1)]
T1 = np.sum(T1, axis=0)
T2 = [np.dot(X[:, i+1:i+2], X[:, i:i+1].T) for i in range(N-1)]
T2 = np.sum(T2, axis=0)
T3 = [np.dot(X[:, i:i+1], X[:, i:i+1].T) for i in range(N-1)]
T3 = np.sum(T3, axis=0)
# T4 = np.dot(X[:, 0:1], X[:, 0:1].T) # this is actually not used
LAM_best = np.dot(T2, np.linalg.inv(T3))
# Kappa_best can be useful for generating samples using (called Sigma in Singh paper)
# x_(n+1) = dot(LAM, x_n) + dot(sqrt(Kappa_best), Xi_n)
Kappa_best = (T1 - np.dot(np.dot(T2, np.linalg.inv(T3)), T2.T)) / N
# J is -lambda in Singh paper
# WARNING: tau is 1 here (if a different one is used J.T gets multiplied by tau)
J = spl.logm(LAM_best)
if not np.all(np.isclose(spl.expm(J), LAM_best, rtol=1e-01)):
print("Warning: logarithm!")
if np.any(np.iscomplex(J)):
J = np.real(J)
print("Warning: complex values in J; casting to real!")
# TODO: implement bayes I for Q0 (called c in Singh paper)
Q0 = T3 / N # this here is bayes II solution (equivalent to sample covariance)
Sigma_best = -np.dot(J, Q0)-np.dot(Q0, J.T)
tau_x = J.diagonal()
np.fill_diagonal(J, 0)
EC_best = J
else:
raise('method should be either \'lyapunov\' or \'moments\'')
self.C = EC_best
self.Sigma = Sigma_best
self.tau_x = tau_x
self.d_fit = d_fit
return self
def obsv(a, c=None):
n, _ = np.shape(a)
c = np.eye(n) if c is None else c
return solve_lyapunov(a.T, -c.T.dot(c))
def ctrl(a, b=None):
n, _ = np.shape(a)
b = np.eye(n) if b is None else b
return solve_lyapunov(a, -b.dot(b.T))
# in which you can specify whether you want left or right handed eigenvectors. # To determine controllability you need to calculate the input pole vectors # which is dependant on the left eigenvectors. # Calculate eigen vectors and pole vectors val, vec = LA.eig(A, None, 1, 0, 0, 0) U1 = np.dot(B.T, vec[:, 0].T) U2 = np.dot(B.T, vec[:, 1].T) # Set-up controllability Matrix Con = np.zeros([2, 2]) Con[:, 0] = B.T Con[:, 1] = (A * B).T n = lin.matrix_rank(Con) P = LA.solve_lyapunov(A, -B * B.T) # Display results print '\nThe transfer function realization is:' print 'G(s) = ' print np.poly1d(G.num[0], variable='s') print "----------------" print np.poly1d(G.den, variable='s') print '\n1) Eigenvalues are: p1 = ', val[0], 'and p2 = ', val[1] print ' with eigenvectors: q1 = ', vec[:, 0], 'and q2 = ', vec[:, 1] print ' Input pole vectors are: up1 = ', U1, 'and up2 = ', U2 print '\n2) The controlabillity matrix has rank', n, 'and is given as:' print ' ', Con[0, :], '\n ', Con[1, :]
# best distance between model and empirical data
best_dist = 1e10
best_Pearson = 0.
# record model parameters and outputs
dist_FC_hist = np.zeros([n_opt]) * np.nan # FC error = matrix distance
Pearson_FC_hist = np.zeros([n_opt]) * np.nan # Pearson corr model/objective
stop_opt = False
i_opt = 0
while not stop_opt:
# calculate Jacobian of dynamical system
J = -np.eye(N) / tau_x + EC
# calculate FC0 and FCtau for model
FC0 = spl.solve_lyapunov(J, -Sigma)
FCtau = np.dot(FC0, spl.expm(J.T * tau))
# calculate error between model and empirical data for FC0 and FC_tau (matrix distance)
err_FC0 = np.sqrt(np.sum((FC0 - FC0_obj)**2)) / np.sqrt(np.sum(FC0_obj**2))
err_FCtau = np.sqrt(np.sum(
(FCtau - FCtau_obj)**2)) / np.sqrt(np.sum(FCtau_obj**2))
dist_FC_hist[i_opt] = 0.5 * (err_FC0 + err_FCtau)
# calculate Pearson corr between model and empirical data for FC0 and FC_tau
Pearson_FC_hist[i_opt] = 0.5 * (
stt.pearsonr(FC0.reshape(-1), FC0_obj.reshape(-1))[0] +
stt.pearsonr(FCtau.reshape(-1), FCtau_obj.reshape(-1))[0])
# best fit given by best Pearson correlation coefficient for both FC0 and FCtau (better than matrix distance)
if dist_FC_hist[i_opt] < best_dist:
def minimize(dynMat,
outMat=None,
structMat=None,
covMat=None,
rankPar=10.,
**optDict):
"""
Minimizes -log(|X|) + rankPar || Z||
subject to dynMat * X + X * dynMat' + Z = 0
structMat .* x - covMat = 0
Outputs:
outDict with keys:
X, Z, Y1, Y2
"""
#-----------------------------------------------------------------
# Initialization and Pre-processing..............................
assert dynMat.ndim == 2
assert covMat is not None, "I need some statistics to run this optimization problem...."
assert covMat.ndim == 2
norm = np.linalg.norm # Makes life easier later
minResPrimal = 1.0e05
# Converting all input arrays into matrices
# IMPORTANT: * NOW REPRESENTS MATRIX MULTIPLICATION AND NOT ELEMENTWISE MULTIPLICATION
warn(
"Be careful with array multiplication. All arrays are now matrices, so A*B is matrix multiplication, not elementwise multiplication."
)
# Dimensionality of the state
nState = dynMat.shape[0]
if outMat is None:
warn("No outMat supplied, using identity")
outMat = np.identity(nState, dtype=dynMat.dtype)
# Take the state as the output if not specified
# Dimensionality of output
nOut = outMat.shape[0]
if structMat is None:
warn(
"No structMat supplied, using non-zero entries of covMat to define one.."
)
# Treat all zero entries of covMat as being unknown.
# It's not ideal, but hey, if you're too lazy to supply a structMat....
structMat = covMat.copy().astype(bool).astype(np.int)
dynMat = np.asmatrix(dynMat)
outMat = np.asmatrix(outMat)
covMat = np.asmatrix(covMat)
structMat = np.asmatrix(structMat)
# Default values for iterative solution of the optimization problem
# Documentation on the iterations to be added soon
stepSize = optDict.get("stepSize", 10.)
tolPrimal = optDict.get("tolPrimal", 1.0e-06)
tolDual = optDict.get("tolDual", 1.0e-06)
iterMax = int(optDict.get("iterMax", 1.0e05))
printIter = int(optDict.get("printIter", 100))
if isinstance(stepSize, np.ndarray): stepSize = stepSize.flatten()[0]
# Initializing X, Z, Y1, and Y2 for the iterations
# Documentation coming soon
Z0 = optDict.get("Z0", None)
X0 = optDict.get("X0", None)
Y10 = optDict.get("Y10", None)
Y20 = optDict.get("Y20", None)
if Z0 is None:
print("No Z0 supplied, using identity...")
Z0 = np.identity(nState, dtype=dynMat.dtype)
if X0 is None:
print("No X0 supplied, solving Lyapunov equation with Z0")
X0 = solve_lyapunov(dynMat, -Z0)
X0 = np.asmatrix(X0)
Z0 = np.asmatrix(Z0)
if Y10 is None:
print("No Y10 supplied, solving adjoing Lyapunov equation")
Y10 = solve_lyapunov(dynMat.H, X0)
if not hasattr(Y10, "H"):
print(
"Y10 from solve_lyapunov does not have attribute 'H', so it's an ndarray and not a matrix. I've recast it into matrix, but have a look at why this is happening...."
)
Y10 = np.asmatrix(Y10)
Y10 = rankPar * Y10 / norm(Y10,
2) # Norm returns largest singular value
else:
Y10 = np.asmatrix(Y10)
# Y2 relates to imposing the constraint due to observed covariance stats,
# so it's size should relate to the dimensionality of the output (not state)
if Y20 is None:
Y20 = np.identity(outMat.shape[0], dtype=dynMat.dtype)
Y20 = np.asmatrix(Y20)
X = X0
Z = Z0
Y1 = Y10
Y2 = Y20
stepSize0 = stepSize
#---------------------------------------------------------------------------------
#---------------------------------------------------------------------------------
# AMA method for optimization:
print(
"Write decorators for timing and logging (have a look at your bookmarks on Chrome)"
)
warn(
"This code currently uses only method AMA. ADMM from the original matlab function hasn't been implemented."
)
# pdb.set_trace()
# Ignoring some lines from original MATLAB code because I don't see why they're needed (yet)
evalsLadjY = np.real( np.linalg.eigvals( dynMat.H * Y1 + Y1*dynMat + \
outMat.H * np.multiply( structMat, Y2) * outMat ) )
logDetLadjY = np.sum(np.log(evalsLadjY + 0.j))
dualY = logDetLadjY - np.trace(covMat * Y2) + nState
#print("Before starting iterations, evalsLadjY, logDetLadjY, dualY, trace(covMat*Y2) are",evalsLadjY, logDetLadjY, dualY, np.trace(covMat*Y2))
Istate = np.identity(nState, dtype=dynMat.dtype) # Identity
beta = 0.7 # Backtracking parameter
stepSize1 = stepSize0 # Initial step size
failIters = [] # Log failed iteration numbers
print("")
print("Starting iterations for AMA......")
print(
"stepSize_BB stepSize tolPrimal resPrimal tolDual abs(dualGap) iter"
)
funPrimalArr = np.zeros(iterMax)
funDualArr = np.zeros(iterMax)
resPrimalArr = np.zeros(iterMax)
dualGapArr = np.zeros(iterMax)
# Will include time later
##=======================================================================================
# Function to dump data
def dumpToFile(dumpFileName, dumpX, dumpOptDict):
if not (dumpFileName.endswith('.h5')
or dumpFileName.endswith('.hdf5')):
dumpFileName = dumpFileName + '.hdf5'
with h5py.File(dumpFileName, "w") as outFile:
dumpData = outFile.create_dataset("X",
data=dumpX,
compression='gzip')
for key in dumpOptDict.keys():
if isinstance(dumpOptDict[key], np.ndarray):
outFile.create_dataset(key,
data=dumpOptDict[key],
compression='gzip')
else:
dumpData.attrs[key] = dumpOptDict[key]
##=======================================================================================
for AMAstep in range(iterMax):
dynMat = np.asmatrix(dynMat)
outMat = np.asmatrix(outMat)
structMat = np.asmatrix(structMat)
Istate = np.asmatrix(Istate)
# Minimization step for X^{k+1}
Xnew = np.linalg.solve(
dynMat.H * Y1 + Y1 * dynMat +
outMat.H * np.multiply(structMat, Y2) * outMat, Istate)
# Xnew = np.linalg.inv( dynMat.H * Y1 + Y1 * dynMat + outMat.H * np.multiply( structMat , Y2) * outMat)
Xnew = np.asmatrix(Xnew)
# To get the inverse of a matrix A, apparently np.linalg.solve(A, Identity) is better than np.linalg.inv(A)
# I'm not sure yet if I need to call this Xnew instead of just X, I'll get back to this later
Xnew = (Xnew + Xnew.H) / 2.
eigX = np.real(np.linalg.eigvals(Xnew))
logDetX = np.sum(np.log(eigX + 0.j))
# Gradient of the dual function
gradD1 = dynMat * Xnew + Xnew * dynMat.H
gradD2 = np.multiply(structMat, outMat * Xnew * outMat.H) - covMat
#print("|Xnew|,|gradD1|, |gradD2|:",det(Xnew),det(gradD1), det(gradD2))
#print()
# Define the primal residual as difference in new statistics from observed (constraints)
resPrimalNew2 = gradD2
stepSize = stepSize1
#if AMAstep == 0:
#print()
#print("evalsX, logDetX, |gradD1|, |gradD2| are",eigX, logDetX, det(gradD1), det(gradD2))
for innerIter in range(100):
a = rankPar / stepSize
aTol = 1.0e-01
if False:
if a > aTol: a = aTol
assert isinstance(
dynMat, np.matrix) and (Xnew, np.matrix) and (Y1, np.matrix)
# Minimization step for Z^{k+1}
Wnew = -(dynMat * Xnew + Xnew * dynMat.H + (1. / stepSize) * Y1)
U, svals, V = np.linalg.svd(Wnew)
warn("numpy's SVD factors as M=USV, not USV*")
# The above line differs from the original matlab code because numpy's svd returns the singular values as an array directly, while Matlab returns a diagonal matrix
if isinstance(svals, np.matrix):
warn(
"The array of singular values is a matrix and not a regular array"
)
svals = svals.A
assert isinstance(U, np.matrix) and isinstance(V, np.matrix)
svals = np.abs(svals)
# Singular value thresholding
if True:
svalsNew = ((1. - a / np.abs(svals)) *
svals) * (np.abs(svals) > a).astype(int)
svalsNew = np.abs(svalsNew)
# So what happened there is, for svals>a, sNew = (1 - a/|s|) * s = s-aA
# That doesn't make much sense to me, but we'll see how it goes
# Singular values should always be positive, so I don't see why all of that extra stuff is needed...
else:
svals = np.abs(svals)
svalsNew = svals.copy()
svalsNew[svals <= a] = 0.
svalsNew[svals > a] = svals - a
#print()
#print("|Wnew|, a,rankPar, stepSize, svalsNew:", det(Wnew), a,rankPar,stepSize,svalsNew)
# Update Z
Znew = U * np.asmatrix(np.diag(svalsNew)) * V
Znew = (Znew + Znew.H) / 2.
# Update Y
resPrimalNew1 = gradD1 + Znew
# Lagrange multiplier update
Y1new = Y1 + stepSize * resPrimalNew1
Y2new = Y2 + stepSize * resPrimalNew2
Y1new = (Y1new + Y1new.H) / 2. # Keep them Hermitian
Y2new = (Y2new + Y2new.H) / 2.
#print("|Znew|,|Y1new|,|Y2new|:", det(Znew), det(Y1new), det(Y2new))
assert isinstance(Y1new, np.matrix) and (Y2new, np.matrix)
# Eigenvalues of X^{-1} (why though?)
evalsLadjYnew = np.real(
np.linalg.eigvals(dynMat.H * Y1new + Y1new * dynMat +
outMat.H * np.multiply(structMat, Y2new) *
outMat))
logDetLadjYnew = np.sum(np.log(evalsLadjYnew + 0.j))
dualYnew = logDetLadjYnew - np.trace(covMat * Y2new) + nState
#print("evals:",evalsLadjYnew)
#print("logdetLadjYnew, dualYnew:", logDetLadjYnew, dualYnew)
if np.amin(evalsLadjYnew) < 0.:
stepSize = stepSize * beta # Backtrack the step
#print("Reducing step size coz evals<0.....")
elif ( dualYnew < dualY + \
np.trace( gradD1 *(Y1new - Y1)) + \
np.trace( gradD2 *(Y2new - Y2)) - \
(0.5/stepSize) * norm(Y1new - Y1, ord='fro')**2 - \
(0.5/stepSize) * norm(Y2new - Y2, ord='fro')**2 ):
# Note: Frobenius norm essentially flattens the array and calculates a vector 2-norm
stepSize = stepSize * beta
#print("Reducing step size coz dualYnew < .....")
else:
#print("Not reducing the step size...")
break
# So this is where we're breaking out of the inner loop
# I suppose what's happening here is we're checking if the primal residual is acceptable
# Primal residual
resPrimalNew1 = dynMat * Xnew + Xnew * dynMat.H + Znew
resPrimalNew2 = np.multiply(structMat,
outMat * Xnew * outMat.H) - covMat
resPrimal = np.sqrt(
norm(resPrimalNew1, ord='fro')**2 +
norm(resPrimalNew2, ord='fro')**2)
#print("resPrimal:",resPrimal)
# Calculating the duality gap
dualGap = -logDetX + rankPar * np.sum(svalsNew) - dualYnew
#print("dualGap:",dualGap)
# Print progress for every 100 outer iterations
#if AMAstep%100 == 0:
if resPrimal < 0.9 * minResPrimal:
minResPrimal = resPrimal
#if AMAstep%printIter == 0:
print("%7.2g %8.3g %6.2g %10.4g %6.2g %7.2g %6.4g %d" %(stepSize1, \
stepSize, tolPrimal, resPrimal, tolDual, (dualGap), np.sum(svalsNew), AMAstep) )
sys.stdout.flush()
#pdb.set_trace()
# We start BB stepsize selection now, apparently, whatever that is..
Xnew1 = np.linalg.solve( \
dynMat.H * Y1new + Y1new * dynMat + outMat.H * np.multiply( structMat, Y2new) * outMat, \
Istate)
Xnew1 = np.asmatrix(Xnew1)
Xnew1 = (Xnew1 + Xnew1.H) / 2.
gradD1new = dynMat * Xnew1 + Xnew1 * dynMat.H
gradD2new = np.multiply(structMat, outMat * Xnew1 * outMat.H) - covMat
gradD1new = np.asmatrix(gradD1new)
gradD2new = np.asmatrix(gradD2new)
Y1new = np.asmatrix(Y1new)
Y2new = np.asmatrix(Y2new)
stepSize1 = np.real( ( norm(Y1new - Y1, ord='fro')**2 + norm(Y2new - Y2, ord='fro')**2 )/ \
( np.trace( (Y1new - Y1) * (gradD1 - gradD1new) ) + np.trace( (Y2new - Y2) * (gradD2 - gradD2new) ) ) )
#print("|Xnew1|, |gradD1new|, |gradD2new|, stepSize1:",det(Xnew1), det(gradD1new), det(gradD2new), stepSize1)
if (stepSize1 < 0.) or not np.isfinite(stepSize1):
stepSize1 = stepSize0
failIters.append(AMAstep)
#sys.exit()
# This is the end of the iteration... Apparently
funPrimalArr[AMAstep] = -np.log(
np.linalg.det(Xnew + 0.j)) + rankPar * norm(Znew, ord='nuc')
funDualArr[AMAstep] = np.real(dualYnew)
resPrimalArr[AMAstep] = resPrimal
dualGapArr[AMAstep] = np.abs(dualGap)
# Again, I'll have to include execution time info later
# Stopping criteria
if (np.abs(dualGap) < tolDual) and (resPrimal < tolPrimal):
print("AMA converged to assigned accuracy!")
print("AMA steps: %d" % AMAstep)
print("%12.2g %10.2g %10.2g %10.2g %10.2g %10.2g %d" %(stepSize1, \
stepSize, tolPrimal, resPrimal, tolDual, np.abs(dualGap), AMAstep) )
break
Y1 = Y1new
Y2 = Y2new
dualY = dualYnew
# Assigning an output dictionary
outDict = {}
if AMAstep == iterMax:
outDict['flag'] = 0
else:
outDict['flag'] = 1
outDict['X'] = Xnew
outDict['Z'] = Znew
outDict['Y1'] = Y1new
outDict['Y2'] = Y2new
outDict['steps'] = AMAstep
outDict['funPrimalArr'] = funPrimalArr[:AMAstep]
outDict['funDualArr'] = funDualArr[:AMAstep]
outDict['dualGapArr'] = dualGapArr[:AMAstep]
return outDict
def solveQ(self):
return solve_lyapunov((transpose(self.W) - self.I), -2 * self.I)
def check_continuous_case(self, a, q):
x = solve_lyapunov(a, q)
assert_array_almost_equal(
np.dot(a, x) + np.dot(x,
a.conj().transpose()), q)
def optimize(FC0_obj,FC1_obj,tau_x,mask_EC,mask_Sigma):
N = FC0_obj.shape[0]
# optimzation rates (to avoid explosion of activity, Sigma is tuned quicker)
epsilon_EC = 0.0005
epsilon_Sigma = 0.05
min_val_EC = 0. # minimal value for tuned EC elements
max_val_EC = 1. # maximal value for tuned EC elements
min_val_Sigma = 0.01 # minimal value for tuned Sigma elements
# initial EC
EC = np.zeros([N,N]) # initial connectivity
Sigma = np.eye(N) # initial noise
# record best fit (matrix distance between model and empirical FC)
best_dist = 1e10
# scaling coefs for FC0 and FC1
a0 = np.linalg.norm(FC1_obj) / (np.linalg.norm(FC0_obj) + np.linalg.norm(FC1_obj))
a1 = 1. - a0
stop_opt = False
i_opt = 0
while not stop_opt:
# calculate Jacobian of dynamical system
J = -np.eye(N)/tau_x + EC
# calculate FC0 and FC1 for model
FC0 = spl.solve_lyapunov(J,-Sigma)
FC1 = np.dot(FC0,spl.expm(J.T))
# matrices of model error
Delta_FC0 = FC0_obj-FC0
Delta_FC1 = FC1_obj-FC1
# calculate error between model and empirical data for FC0 and FC_tau (matrix distance)
dist_FC_tmp = 0.5*(np.linalg.norm(Delta_FC0)/np.linalg.norm(FC0_obj)+np.linalg.norm(Delta_FC1)/np.linalg.norm(FC1_obj))
# calculate Pearson correlation between model and empirical data for FC0 and FC_tau
Pearson_FC_tmp = 0.5*(stt.pearsonr(FC0.reshape(-1),FC0_obj.reshape(-1))[0]+stt.pearsonr(FC1.reshape(-1),FC1_obj.reshape(-1))[0])
# record best model parameters
if dist_FC_tmp<best_dist:
best_dist = dist_FC_tmp
best_Pearson = Pearson_FC_tmp
i_best = i_opt
J_mod_best = np.array(J)
Sigma_mod_best = np.array(Sigma)
else:
stop_opt = i_opt>50
# Jacobian update
Delta_J = np.dot(np.linalg.pinv(FC0),a0*Delta_FC0+np.dot(a1*Delta_FC1,spl.expm(-J.T))).T
# update EC (recurrent connectivity)
EC[mask_EC] += epsilon_EC * Delta_J[mask_EC]
EC[mask_EC] = np.clip(EC[mask_EC],min_val_EC,max_val_EC)
# update Sigma (input variances)
Delta_Sigma = -np.dot(J,Delta_FC0)-np.dot(Delta_FC0,J.T)
Sigma[mask_Sigma] += epsilon_Sigma * Delta_Sigma[mask_Sigma]
Sigma[mask_Sigma] = np.maximum(Sigma[mask_Sigma],min_val_Sigma)
# check for stop
if not stop_opt:
i_opt += 1
else:
print('stop at step',i_best,'with best FC dist:',best_dist,'; best FC Pearson:',best_Pearson)
return (J_mod_best,Sigma_mod_best)
def _normtail(sig, A, x, C):
# observability gramiam when A perturbed by sig
W = solve_lyapunov(A.T + sig*np.eye(len(A)), -C.T.dot(C))
return np.sqrt(x.dot(W).dot(x.T) / 2 / sig) # eq (39)
def matt_mou(ts_emp, mask_EC=None, verbose=0, true_C=None, true_S=None):
# optimzation steps and rate
n_opt = 10000
epsilon_EC = 0.0005
epsilon_Sigma = 0.1
regul_EC = 0 # 0.5
regul_Sigma = 0 # 0.001
print('regularization:', regul_EC, ';', regul_Sigma)
N = 50 # number of ROIs
n_tau = 3 # number of time shifts for FC_emp
v_tau = np.arange(n_tau)
i_tau_opt = 1 # time shift for optimization
min_val_EC = 0. # minimal value for EC
max_val_EC = 0.2 # maximal value for EC
min_val_Sigma_diag = 0. # minimal value for Sigma
# FC matrix (ts_emp matrix with stucture time x ROI index)
n_T = ts_emp.shape[0]
ts_emp -= np.outer(np.ones(n_T), ts_emp.mean(0))
FC_emp = np.zeros([n_tau, N, N])
for i_tau in range(n_tau):
FC_emp[i_tau, :, :] = np.tensordot(
ts_emp[0:n_T - n_tau, :],
ts_emp[i_tau:n_T - n_tau + i_tau, :],
axes=(0, 0)) / float(n_T - n_tau - 1)
# autocovariance time constant
log_ac = np.log(np.maximum(FC_emp.diagonal(axis1=1, axis2=2), 1e-10))
lin_reg = np.polyfit(np.repeat(v_tau, N), log_ac.reshape(-1), 1)
tau_x = -1. / lin_reg[0]
print('inverse of negative slope (time constant):', tau_x)
# mask for existing connections for EC and Sigma
mask_diag = np.eye(N, dtype=bool)
if mask_EC is None:
mask_EC = np.logical_not(
mask_diag) # all possible connections except self
mask_Sigma = np.eye(N, dtype=bool) # independent noise
# mask_Sigma = np.ones([N,N],dtype=bool) # coloured noise
# optimization
print('*opt*')
print('i tau opt:', i_tau_opt)
tau = v_tau[i_tau_opt]
# objective FC matrices (empirical)
FC0_obj = FC_emp[0, :, :]
FCtau_obj = FC_emp[i_tau_opt, :, :]
coef_0 = np.sqrt(np.sum(FCtau_obj**2)) / (np.sqrt(np.sum(FC0_obj**2)) +
np.sqrt(np.sum(FCtau_obj**2)))
coef_tau = 1. - coef_0
# initial network parameters
EC = np.zeros([N, N])
Sigma = np.eye(N) # initial noise
# best distance between model and empirical data
best_dist = 1e10
best_Pearson = 0.
# record model parameters and outputs
dist_FC_hist = np.zeros([n_opt]) * np.nan # FC error = matrix distance
Pearson_FC_hist = np.zeros([n_opt
]) * np.nan # Pearson corr model/objective
dist_C_hist = np.zeros([n_opt]) * np.nan # FC error = matrix distance
Pearson_C_hist = np.zeros([n_opt]) * np.nan # Pearson corr model/objective
stop_opt = False
i_opt = 0
while not stop_opt:
# calculate Jacobian of dynamical system
J = -np.eye(N) / tau_x + EC
# calculate FC0 and FCtau for model
FC0 = spl.solve_lyapunov(J, -Sigma)
FCtau = np.dot(FC0, spl.expm(J.T * tau))
# calculate error between model and empirical data for FC0 and FC_tau (matrix distance)
err_FC0 = np.sqrt(np.sum(
(FC0 - FC0_obj)**2)) / np.sqrt(np.sum(FC0_obj**2))
err_FCtau = np.sqrt(np.sum(
(FCtau - FCtau_obj)**2)) / np.sqrt(np.sum(FCtau_obj**2))
dist_FC_hist[i_opt] = 0.5 * (err_FC0 + err_FCtau)
# calculate Pearson corr between model and empirical data for FC0 and FC_tau
Pearson_FC_hist[i_opt] = 0.5 * (
stt.pearsonr(FC0.reshape(-1), FC0_obj.reshape(-1))[0] +
stt.pearsonr(FCtau.reshape(-1), FCtau_obj.reshape(-1))[0])
# calculate the error between true and estimated connectivity C if the true is provided (for debugging)
if not (true_C is None):
dist_C_hist[i_opt] = np.sqrt(
((true_C - EC)**2).sum() / (true_C**2).sum())
Pearson_C_hist[i_opt] = stt.pearsonr(true_C.flatten(),
EC.flatten())[0]
# best fit given by best Pearson correlation coefficient for both FC0 and FCtau (better than matrix distance)
if dist_FC_hist[i_opt] < best_dist:
best_dist = dist_FC_hist[i_opt]
best_Pearson = Pearson_FC_hist[i_opt]
i_best = i_opt
EC_best = np.array(EC)
Sigma_best = np.array(Sigma)
FC0_best = np.array(FC0)
FCtau_best = np.array(FCtau)
else:
stop_opt = i_opt > 100
# Jacobian update with weighted FC updates depending on respective error
Delta_FC0 = (FC0_obj - FC0) * coef_0
Delta_FCtau = (FCtau_obj - FCtau) * coef_tau
Delta_J = np.dot(np.linalg.pinv(FC0), Delta_FC0 +
np.dot(Delta_FCtau, spl.expm(-J.T * tau))).T / tau
# update conectivity and noise
EC[mask_EC] += epsilon_EC * (Delta_J - regul_EC * EC)[mask_EC]
EC[mask_EC] = np.clip(EC[mask_EC], min_val_EC, max_val_EC)
Sigma[mask_Sigma] += epsilon_Sigma * (-np.dot(J, Delta_FC0) - np.dot(
Delta_FC0, J.T) - regul_Sigma)[mask_Sigma]
Sigma[mask_diag] = np.maximum(Sigma[mask_diag], min_val_Sigma_diag)
# check if end optimization: if FC error becomes too large
if stop_opt or i_opt == n_opt - 1:
stop_opt = True
print('stop at step', i_opt, 'with best dist', best_dist,
';best FC Pearson:', best_Pearson)
else:
if (i_opt) % 20 == 0:
print('opt step:', i_opt)
print('current dist FC:', dist_FC_hist[i_opt],
'; current Pearson FC:', Pearson_FC_hist[i_opt])
i_opt += 1
if verbose == 2:
plt.figure()
plt.subplot(121)
plt.plot(dist_FC_hist, label=r'$Q_{0/\tau}$ error')
if not (true_C is None):
plt.plot(dist_C_hist, label='C error')
plt.legend()
plt.ylabel('matrix error')
plt.xlabel('iterations')
plt.subplot(122)
plt.plot(Pearson_FC_hist, label=r'$Q_{0/\tau}$ similarity')
if not (true_C is None):
plt.plot(Pearson_C_hist, label='C similarity')
plt.legend()
plt.ylabel('pearson correlation')
plt.xlabel('iterations')
return EC_best, Sigma_best
we will create the dynamics of a spring-mass-damper system
m \dot\dot x = F - k x - c \dot x
"""
GLOBALS.k = 100.
GLOBALS.c = 5.
GLOBALS.m = 1.
k, c, m = GLOBALS.k, GLOBALS.c, GLOBALS.m
A = np.array([[0., 1.], [-k / m, -c / m]])
b = np.array([0., 1. / m])
Q = -np.eye(2)
P = splin.solve_lyapunov(A.transpose(), Q)
#P = np.dot(X,X.transpose())
def UVtrajectory(t, u0, v0):
X0 = np.array([u0, v0])
Phi_t = splin.expm(t * A)
return np.dot(Phi_t, X0)
if __name__ == '__main__':
plt.close('all')
if False:
xs = np.linspace(-2, 2, 100)
def grammian2(self):
A, B, C = self.state_space()
Q = -B @ B.T
W = solve_lyapunov(A.toarray(), Q.toarray())
return W
def check_continuous_case(self, a, q):
x = solve_lyapunov(a, q)
assert_array_almost_equal(np.dot(a, x) + np.dot(x, a.conj().transpose()), q)
def compute_lyapunov_solution(b):
rhs = np.zeros(b.shape)
rhs[0, 0] = 1.
z_i = la.solve_lyapunov(b, rhs)
return z_i