Welcome to ConIII’s documentation!

coniii.enumerate module

coniii.enumerate.fast_logsumexp(X, coeffs=None)

Simplified version of logsumexp to do correlation calculation in Ising equation files. Scipy’s logsumexp can be around 10x slower in comparison.

X : ndarray
Terms inside logs.
coeffs : ndarray
Factors in front of exponentials.
float
Value of magnitude of quantity inside log (the sum of exponentials).
float
Sign.
coniii.enumerate.get_3idx(n)

Get binary 3D matrix with truth values where index values correspond to the index of all possible ijk parameters. We can do this by recognizing that the pattern along each plane in the third dimension is like the upper triangle pattern that just moves up and over by one block each cut lower into the box.

coniii.enumerate.get_nidx(k, n)

Get the kth order indices corresponding to all the states in which k elements are firing up out of n spins. The ordering correspond to that returned by bin_states().

print where(exact.get_3idx(4)) print where(exact.get_nidx(3,4)) <<<<<

coniii.enumerate.get_terms(subix, prefix, binstate, br, ix0)

Spins are put in explicitly

coniii.enumerate.get_terms01(subix, prefix, binstate, br, ix0)

Specific to {0,1}.

coniii.enumerate.get_terms11(subix, prefix, binstate, br, ix0)

Specific to {-1,1}.

coniii.enumerate.mp_fast_logsumexp(X, coeffs=None)

fast_logsumexp for high precision numbers using mpmath.

X : ndarray
Terms inside logs.
coeffs : ndarray
Factors in front of exponentials.
float
Value of magnitude of quantity inside log (the sum of exponentials).
float
Sign.
coniii.enumerate.pairwise(n, sym=0, **kwargs)

Wrapper for writing pairwise maxent model (Ising) files.

n : int
System size.
sym : int, 0
Can be 0 or 1.

**kwargs

None

coniii.enumerate.triplet(n, sym=0, **kwargs)

Wrapper for writing triplet-order maxent model.

n : int
System size.
sym : int, 0
Can be 0 or 1.

**kwargs

None

coniii.enumerate.write_eqns(n, sym, corrTermsIx, suffix='', high_prec=False)

Create strings for writing out the equations and then write them to file.

TODO: This code needs some cleanup.

n : int
number of spins
sym : int
value of 1 will use {-1,1} formulation, 0 means {0,1}
corrTermsIx : list of ndarrays
Allows specification of arbitrary correlations to constrain using an index based structure. These should be index arrays as would be returned by np.where that specify which correlations to write down. Each consecutive array should specify a matrix of sequentially increasing dimension. [Nx1, NxN, NxNxN, …]

suffix : str, ‘’ high_prec : bool, False

coniii.enumerate.write_py(n, sym, contraintTermsIx, signs, expterms, Z, extra='', suffix='', high_prec=False)

Write out Ising equations for Python.

n : int
System size.

contraintTermsIx : list of str signs : list of ndarray

Sign for each term in the numerator when computing correlations.
expterms : list of str
Every single energy term.
Z : str
Energies for all states that will be put into partition function.
extra : str, ‘’
any extra lines to add at the end

suffix : str, ‘’ high_prec : bool, False

If True, write version that uses mpmath for high precision calculations.

coniii.enumerate_potts module

class coniii.enumerate_potts.PythonFileWriterBase(n, k)

Bases: object

energy_terms_generator()

Generator for iterating through all possible states and yield the energy expression as well as the configuration of spins.

write(fname)

Write equations to file.

class coniii.enumerate_potts.SpecificFieldGenericCouplings(n, k)

Bases: coniii.enumerate_potts.PythonFileWriterBase

This version specifies a field for every distinct Potts state, but considers correlation averaged over all possible states as long as they agree.

When writing the equations, the fields are assumed to first increase by spin index then by state index (fields for the n spins for k=0 come first, then for k=1, etc.). The couplings come after in the conventional order of ij where j increases up to i before i is incremented.

energy_terms_generator()

Generator for iterating through all possible states and yield the energy expression as well as the configuration of spins. The energy expression is returned as a string and assumes that the field come first followed by the couplings.

coniii.enumerate_potts.insert_newlines(s, n)

Insert character every n in list.

coniii.enumerate_potts.split_string(s, n)

Insert character every n.

coniii.ising package

Submodules

coniii.ising.automaton module

class coniii.ising.automaton.Ising2D(dim, J, h=0, rng=None)

Bases: object

Simulation of the ferromagnetic Ising model on a 2D periodic lattice with quenched disorder in the local fields.

flip_metropolis

Flip a single lattice spin using Metropolis sampling.

i : int j : int

iterate(n_iters, systematic=True)

n_iters : int systematic : bool,True

If True, iterate through each spin on the lattice in sequence.
coniii.ising.automaton.coarse_grain(lattice, factor)

Block spin renormalization with majority rule.

lattice : ndarray
+/-1

factor : int

renormalized_lattice : ndarray

coniii.ising.test_automaton module

Module contents

coniii.solvers module

class coniii.solvers.ClusterExpansion(sample, model=None, calc_observables=None, sample_size=1000, iprint=True, **default_model_kwargs)

Bases: coniii.solvers.Solver

Implementation of Adaptive Cluster Expansion for solving the inverse Ising problem, as described in John Barton and Simona Cocco, J. of Stat. Mech. P03002 (2013).

Specific to pairwise Ising constraints.

S(cluster, coocMat, deltaJdict={}, useAnalyticResults=False, priorLmbda=0.0, numSamples=None)

Calculate pairwise entropy of cluster. (First fits pairwise Ising model.)

cluster : list
List of indices belonging to each cluster.
coocMat : ndarray
Pairwise correlations.

deltaJdict : dict, {} useAnalyticResults : bool, False

Probably want False until analytic formulas are changed to include prior on J

entropy : float Jfull : ndarray

Matrix of couplings.
Sindependent(cluster, coocMat)

Entropy approximation assuming that each cluster appears independently of the others.

cluster : list coocMat : ndarray

Pairwise correlations.
float
Sind, independent entropy.
ndarray
Pairwise couplings.
clusterID(cluster)
deltaS(cluster, coocMat, deltaSdict=None, deltaJdict=None, iprint=True, meanFieldRef=False, priorLmbda=0.0, numSamples=None, independentRef=False, meanFieldPriorLmbda=None)
cluster : list
List of indices in cluster

coocMat : ndarray deltaSdict : dict, None deltaJdict : dict, None iprint : bool, True meanFieldRef : bool, False numSamples : int, None independentRef : bool, False

If True, expand about independent entropy
meanFieldRef : bool, False
If True, expand about mean field entropy
float
deltaScluster
float
deltaJcluster
solve(threshold, cluster=None, deltaSdict=None, deltaJdict=None, iprint=True, priorLmbda=0.0, numSamples=None, meanFieldRef=False, independentRef=True, veryVerbose=False, meanFieldPriorLmbda=None, full_output=False)

threshold : float meanFieldRef : bool, False

Expand about mean-field reference.
independentRef : bool, True
Expand about independent reference.
priorLmbda : float, 0.
Strength of non-interacting prior.
meanFieldPriorLmbda : float, None
Strength of non-interacting prior in mean field calculation (defaults to priorLmbda).
ndarray
Solved multipliers (parameters). For Ising problem, these can be converted into matrix format using utils.vec2mat.
float (optional, only if full_output=True)
Estimated entropy.
ndarray
Solved multipliers (parameters). For Ising problem, these can be converted into matrix format using utils.vec2mat.
list (optional, only if full_output=True)
List of clusters.
dict (optional, only if full_output=True)
deltaSdict
dict (optional, only if full_output=True)
deltaJdict
subsets(thisSet, size, sort=False)

Given a list, returns a list of all unique subsets of that list with given size.

thisSet : list size : int sort : bool, False

list
All subsets of given size.
class coniii.solvers.Enumerate(sample=None, model=None, calc_observables=None, iprint=True, **default_model_kwargs)

Bases: coniii.solvers.Solver

Class for solving fully-connected inverse Ising model problem by enumeration of the partition function and then using gradient descent.

solve(initial_guess=None, constraints=None, max_param_value=50, full_output=False, use_root=True, scipy_solver_kwargs={'method': 'krylov', 'options': {'fatol': 1e-13, 'xatol': 1e-13}})

Must specify either constraints (the correlations) or samples from which the correlations will be calculated using self.calc_observables. This routine by default uses scipy.optimize.root to find the solution. This is MUCH faster than the scipy.optimize.minimize routine which can be used instead.

If still too slow, try adjusting the accuracy.

If not converging, try increasing the max number of iterations.

If receiving Jacobian error (or some other numerical estimation error), parameter values may be too large for faithful evaluation. Try decreasing max_param_value.

initial_guess : ndarray, None
Initial starting guess for parameters. By default, this will start with all zeros if left unspecified.
constraints : ndarray, None
Can specify constraints directly instead of using the ones calculated from the sample. This can be useful when the pairwise correlations are known exactly. This will override the self.constraints data member.
max_param_value : float, 50
Absolute value of max parameter value. Bounds can also be set in the kwargs passed to the minimizer, in which case this should be set to None.
full_output : bool, False
If True, return output from scipy.optimize.minimize.
use_root : bool, True
If False, use scipy.optimize.minimize instead. This is typically much slower.
scipy_solver_kwargs : dict, {‘method’:’krylov’, ‘options’:{‘fatol’:1e-13,’xatol’:1e-13}}
High accuracy is slower. Although default accuracy may not be so good, lowering these custom presets will speed things up. Choice of the root finding method can also change runtime and whether a solution is found or not. Recommend playing around with different solvers and tolerances or getting a close approximation using a different method if solution is hard to find.
ndarray
Solved multipliers (parameters). For Ising problem, these can be converted into matrix format using utils.vec2mat.
dict, optional
Output from scipy.optimize.root.
class coniii.solvers.MCH(sample, model=None, calc_observables=None, sample_size=1000, sample_method='metropolis', mch_approximation=None, iprint=True, sampler_kw={}, **default_model_kwargs)

Bases: coniii.solvers.Solver

Class for solving maxent problems using the Monte Carlo Histogram method.

Broderick, T., Dudik, M., Tkacik, G., Schapire, R. E. & Bialek, W. Faster solutions of the inverse pairwise Ising problem. arXiv 1-8 (2007).

estimate_jac(eps=0.001)

Approximation Jacobian using the MCH approximation.

eps : float, 1e-3

jac : ndarray
Jacobian is an n x n matrix where each row corresponds to the behavior of fvec wrt to a single parameter.
learn_parameters_mch(estConstraints, constraints, maxdlamda=1, maxdlamdaNorm=1, maxLearningSteps=50, eta=1)
estConstraints : ndarray
Constraints estimated from MCH approximation.

constraints : ndarray maxdlamda : float, 1

Max allowed magnitude for any element of dlamda vector before exiting.
maxdlamdaNorm : float, 1
Max allowed norm of dlamda vector before exiting.
maxLearningSteps : int
max learning steps before ending MCH
eta : float, 1
factor for changing dlamda
ndarray
MCH estimate for constraints from parameters lamda+dlamda.
solve(initial_guess=None, constraints=None, tol=None, tolNorm=None, n_iters=30, burn_in=30, maxiter=10, custom_convergence_f=None, iprint=False, full_output=False, learn_params_kwargs={'eta': 1, 'maxdlamda': 1}, generate_kwargs={})

Solve for maxent model parameters using MCH routine.

initial_guess : ndarray, None
Initial starting point.
constraints : ndarray, None
For debugging! Vector of correlations to fit.
tol : float, None
Maximum error allowed in any observable.
tolNorm : float, None
Norm error allowed in found solution.
n_iters : int, 30
Number of iterations to make between samples in MCMC sampling.
burn_in : int, 30
Initial burn in from random sample when MC sampling.
max_iter : int, 10
Max number of iterations of MC sampling and MCH approximation.
custom_convergence_f : function, None

Function for determining convergence criterion. At each iteration, this function should return the next set of learn_params_kwargs and optionally the sample size.

As an example: def learn_settings(i):

‘’’ Take in the iteration counter and set the maximum change allowed in any given parameter (maxdlamda) and the multiplicative factor eta, where d(parameter) = (error in observable) * eta.

Additional option is to also return the sample size for that step by returning a tuple. Larger sample sizes are necessary for higher accuracy. ‘’’ if i<10:

return {‘maxdlamda’:1,’eta’:1}
else:
return {‘maxdlamda’:.05,’eta’:.05}

iprint : bool, False full_output : bool, False

If True, also return the errflag and error history.

learn_parameters_kwargs : dict, {‘maxdlamda’:1,’eta’:1} generate_kwargs : dict, {}

ndarray
Solved multipliers (parameters). For Ising problem, these can be converted into matrix format using utils.vec2mat.
int
Error flag. 0, converged within given criterion 1, max iterations reached
ndarray
Log of errors in matching constraints at each step of iteration.
class coniii.solvers.MCHIncompleteData(*args, **kwargs)

Bases: coniii.solvers.MCH

Class for solving maxent problems using the Monte Carlo Histogram method on incomplete data where some spins may not be visible.

Broderick, T., Dudik, M., Tkacik, G., Schapire, R. E. & Bialek, W. Faster solutions of the inverse pairwise Ising problem. arXiv 1-8 (2007).

NOTE: This only works for Ising model.
Not ready for release.
generate_samples(n_iters, burn_in, uIncompleteStates=None, f_cond_sample_size=None, f_cond_sample_iters=None, sample_size=None, sample_method=None, initial_sample=None, run_regular_sampler=True, run_cond_sampler=True, disp=0, generate_kwargs={})

Wrapper around generate_samples_parallel() from available samplers.

n_iters : int burn_in : int

I think burn in is handled automatically in REMC.

uIncompleteStates : list of unique states f_cond_sample_size : lambda function

Given the number of hidden spins, return the number of samples to take.
f_cond_sample_iters : lambda function
Given the number of hidden spins, return the number of MC iterations to make.

sample_size : int sample_method : str initial_sample : ndarray generate_kwargs : dict

learn_parameters_mch(estConstraints, fullFraction, uIncompleteStates, uIncompleteStatesCount, maxdlamda=1, maxdlamdaNorm=1, maxLearningSteps=50, eta=1)

Update parameters with MCH step. Update is proportional to the difference between the observables and the predicted observables after a small change to the parameters. This is calculated from likelihood maximization, and for the incomplete data points this corresponds to the marginal probability distribution weighted with the number of corresponding data points.

estConstraints : ndarray fullFraction : float

Fraction of data points that are complete.
uIncompleteStates : list-like
Unique incomplete states in data.
uIncompleteStatesCount : list-like
Frequency of each unique data point.

maxdlamda : float,1 maxdlamdaNorm : float,1 maxLearningSteps : int

max learning steps before ending MCH
eta : float,1
factor for changing dlamda

estimatedConstraints : ndarray

solve(X=None, constraints=None, initial_guess=None, cond_sample_size=100, cond_sample_iters=100, tol=None, tolNorm=None, n_iters=30, burn_in=30, maxiter=10, disp=False, full_output=False, learn_params_kwargs={}, generate_kwargs={})

Solve for parameters using MCH routine.

X : ndarray constraints : ndarray

Constraints calculated from the incomplete data (accounting for missing data points).
initial_guess : ndarray=None
initial starting point
cond_sample_size : int or function
Number of samples to make for conditional distribution. If function is passed in, it will be passed number of missing spins and must return an int.
cond_sample_iters : int or function
Number of MC iterations to make between samples.
tol : float=None
maximum error allowed in any observable
tolNorm : float
norm error allowed in found solution
n_iters : int=30
Number of iterations to make between samples in MCMC sampling.

burn_in (int=30) disp : int=0

0, no output 1, some detail 2, most detail
full_output : bool,False
Return errflag and errors at each iteration if True.

learn_parameters_kwargs : dict generate_kwargs : dict

parameters : ndarray
Found solution.

errflag : int errors : ndarray

Errors in matching constraints at each step of iteration.
class coniii.solvers.MPF(sample, model=None, calc_observables=None, calc_de=None, adj=None, iprint=True, **default_model_kwargs)

Bases: coniii.solvers.Solver

K(Xuniq, Xcount, adjacentStates, params)

Compute objective function.

Xuniq : ndarray
(ndata x ndims) unique states that appear in the data
Xcount : ndarray of int
number of times that each unique state appears in the data
adjacentStates : list of ndarray
list of adjacent states for each given unique state
params : ndarray
parameters for computation of energy

K : float

list_adjacent_states(Xuniq, all_connected)

Use self.adj to evaluate all adjacent states in Xuniq.

Xuniq : ndarray all_connected : bool

adjacentStates

logK(Xuniq, Xcount, adjacentStates, params)

Compute log of objective function.

Xuniq : ndarray
(n_samples, n_dim) unique states that appear in the data
Xcount : ndarray of int
number of times that each unique state appears in the data
adjacentStates : list of ndarray
list of adjacent states for each given unique state
params : ndarray
parameters for computation of energy

logK : float

solve(initial_guess=None, method='L-BFGS-B', full_output=False, all_connected=True, parameter_limits=100, solver_kwargs={'disp': False, 'ftol': 1e-15, 'maxiter': 100}, uselog=True)

Minimize MPF objective function using scipy.optimize.minimize.

initial_guess : ndarray, None method : str, ‘L-BFGS-B’

Option for scipy.optimize.minimize.

full_output : bool, False all_connected : bool, True

Switch for summing over all states that data sets could be connected to or just summing over non-data states (second summation in Eq 10 in Sohl-Dickstein 2011).
parameter_limits : float, 100
Maximum allowed magnitude of any single parameter.
solver_kwargs : dict, {‘maxiter’:100,’disp’:False,’ftol’:1e-15}
For scipy.optimize.minimize.
uselog : bool, True
If True, calculate log of the objective function. This can help with numerical precision errors.
ndarray
Solved multipliers (parameters). For Ising problem, these can be converted into matrix format using utils.vec2mat.
dict (optional)
Output from scipy.optimize.minimize returned if full_output is True.
static worker_objective_task(s, Xcount, adjacentStates, params, calc_e)
coniii.solvers.MonteCarloHistogram

alias of coniii.solvers.MCH

class coniii.solvers.Pseudo(sample, model=None, calc_observables=None, get_multipliers_r=None, calc_observables_r=None, k=2, iprint=True, **default_model_kwargs)

Bases: coniii.solvers.Solver

Pseudolikelihood approximation to solving the inverse Ising problem as described in Aurell and Ekeberg, PRL 108, 090201 (2012).

cond_hess(r, X, Jr, pairCoocRhat=None)

Returns d^2 cond_log_likelihood / d Jri d Jrj, with shape (dimension of system)x(dimension of system)

Current implementation uses more memory for speed. For large sample size, it may make sense to break up differently if too much memory is being used.

Deprecated.

pairCooc : ndarray, None
Pass pair_cooc_mat(X) to speed calculation.
cond_jac(r, X, Jr)

Returns d cond_log_likelihood / d Jr, with shape (dimension of system)

Deprecated.

cond_log_likelihood(r, X, Jr)

Equals the conditional log likelihood -L_r.

Deprecated.

r : int
individual index
X : ndarray
binary matrix, (# X) x (dimension of system)
Jr : ndarray
(dimension of system) x (1)

float

pair_cooc_mat(X)

Returns matrix of shape (self.n)x(# X)x(self.n).

For use with cond_hess.

Slow because I haven’t thought of a better way of doing it yet.

Deprecated.

pseudo_log_likelihood(X, J)

TODO: Could probably be made more efficient.

Deprecated.

X : ndarray
binary matrix, (# of samples) x (dimension of system)
J : ndarray
(dimension of system) x (dimension of system) J should be symmetric
solve(force_general=False, **kwargs)

Uses a general all-purpose optimization to solve the problem using functions defined in self.get_multipliers_r and self.calc_observables_r.

force_general : bool, False
If True, force use of “general” algorithm.
initial_guess : ndarray, None
Initial guess for the parameter values.
solver_kwargs : dict, {}
kwargs for scipy.minimize().
ndarray
Solved multipliers (parameters). For Ising problem, these can be converted into matrix format using utils.vec2mat.
class coniii.solvers.RegularizedMeanField(sample, model=None, calc_observables=None, sample_size=1000, iprint=False, **default_model_kwargs)

Bases: coniii.solvers.Solver

Implementation of regularized mean field method for solving the inverse Ising problem, as described in Daniels, Bryan C., David C. Krakauer, and Jessica C. Flack. ``Control of Finite Critical Behaviour in a Small-Scale Social System.’’ Nature Communications 8 (2017): 14301. doi:10.1038/ncomms14301

Specific to pairwise Ising constraints.

bracket1d(xList, funcList)

Assumes xList is monotonically increasing

Get bracketed interval (a,b,c) with a < b < c, and f(b) < f(a) and f(c). (Choose b and c to make f(b) and f(c) as small as possible.)

If minimum is at one end, raise error.

solve(n_grid_points=200, min_size=0, reset_rng=True, min_covariance=False, min_independent=True, cooc_cov=None, priorLmbda=0.0, bracket=None)

Varies the strength of regularization on the mean field J to best fit given cooccurrence data.

n_grid_points : int, 200
If bracket is given, first test at n_grid_points points evenly spaced in the bracket interval, then give the lowest three points to scipy.optimize.minimize_scalar
min_size : int, 0
Use a modified model in which samples with fewer ones than min_size are not allowed.
reset_rng: bool, True
Reset random number generator seed before sampling to ensure that objective function does not depend on generator state.
min_covariance : bool, False
** As of v1.0.3, not currently supported ** Minimize covariance from emperical frequencies (see notes); trying to avoid biases, as inspired by footnote 12 in TkaSchBer06
min_independent : bool, True
** As of v1.0.3, min_independent is the only mode currently supported ** Each <xi> and <xi xj> residual is treated as independent
cooc_cov : ndarray, None
** As of v1.0.3, not currently supported ** Provide a covariance matrix for residuals. Should typically be coocSampleCovariance(samples). Only used if min_covariance and min_independent are False.
priorLmbda : float,0.
** As of v1.0.3, not currently implemented ** Strength of noninteracting prior.
ndarray
Solved multipliers (parameters). For Ising problem, these can be converted into matrix format using utils.vec2mat.
class coniii.solvers.Solver

Bases: object

Base class for declaring common methods and attributes for inverse maxent algorithms.

basic_setup(sample_or_n=None, model=None, calc_observables=None, iprint=True, model_kwargs={})

General routine for setting up a Solver instance.

sample_or_n : ndarray or int, None

If ndarray, of dimensions (samples, dimension).

If int, specifies system size.

If None, many of the default class members cannot be set and then must be set manually.

model : class like one from models.py, None
By default, will be set to solve Ising model.
calc_observables : function, None
For calculating observables from a set of samples.
iprint : str, True
If empty, do not display warning messages.
model_kwargs : dict, {}
Additional arguments that will be passed to Ising class. These only matter if model is None. Important ones include “n_cpus” and “rng”.
fill_in(x, fill_value=0)

Helper function for filling in missing parameter values.

x : ndarray fill_value : float, 0

ndarray
With missing entries filled in.
set_insertion_ix()

Calculate indices to fill in with zeros to “fool” code that takes full set of params.

solve()

To be defined in derivative classes.

class coniii.solvers.SparseEnumerate(sample=None, model=None, calc_observables=None, parameter_ix=None, iprint=True, **default_model_kwargs)

Bases: coniii.solvers.Solver

Class for solving Ising model with a sparse parameter set by enumeration of the partition function and then using gradient descent. Unspecified parameters are implicitly fixed to be zero, which corresponds to leaving the corresponding correlation function unconstrained.

solve(initial_guess=None, constraints=None, max_param_value=50, full_output=False, use_root=True, scipy_solver_kwargs={'method': 'krylov', 'options': {'fatol': 1e-13, 'xatol': 1e-13}})

Must specify either constraints (the correlations) or samples from which the correlations will be calculated using self.calc_observables. This routine by default uses scipy.optimize.root to find the solution. This is MUCH faster than the scipy.optimize.minimize routine which can be used instead.

If still too slow, try adjusting the accuracy.

If not converging, try increasing the max number of iterations.

If receiving Jacobian error (or some other numerical estimation error), parameter values may be too large for faithful evaluation. Try decreasing max_param_value.

initial_guess : ndarray, None
Initial starting guess for parameters. By default, this will start with all zeros if left unspecified.
constraints : ndarray, None
Can specify constraints directly instead of using the ones calculated from the sample. This can be useful when the pairwise correlations are known exactly. This will override the self.constraints data member.
max_param_value : float, 50
Absolute value of max parameter value. Bounds can also be set in the kwargs passed to the minimizer, in which case this should be set to None.
full_output : bool, False
If True, return output from scipy.optimize.minimize.
use_root : bool, True
If False, use scipy.optimize.minimize instead. This is typically much slower.
scipy_solver_kwargs : dict, {‘method’:’krylov’, ‘options’:{‘fatol’:1e-13,’xatol’:1e-13}}
High accuracy is slower. Although default accuracy may not be so good, lowering these custom presets will speed things up. Choice of the root finding method can also change runtime and whether a solution is found or not. Recommend playing around with different solvers and tolerances or getting a close approximation using a different method if solution is hard to find.
ndarray
Solved multipliers (parameters).
dict, optional
Output from scipy.optimize.root.
class coniii.solvers.SparseMCH(sample, model=None, calc_observables=None, sample_size=1000, sample_method='metropolis', mch_approximation=None, parameter_ix=None, iprint=True, sampler_kw={}, **default_model_kwargs)

Bases: coniii.solvers.Solver

Class for solving maxent problems on sparse constraints using the Monte Carlo Histogram method.

See MCH class.

learn_parameters_mch(estConstraints, constraints, maxdlamda=1, maxdlamdaNorm=1, maxLearningSteps=50, eta=1)
estConstraints : ndarray
Constraints estimated from MCH approximation.

constraints : ndarray maxdlamda : float, 1

Max allowed magnitude for any element of dlamda vector before exiting.
maxdlamdaNorm : float, 1
Max allowed norm of dlamda vector before exiting.
maxLearningSteps : int
max learning steps before ending MCH
eta : float, 1
factor for changing dlamda
ndarray
MCH estimate for constraints from parameters lamda+dlamda.
solve(initial_guess=None, constraints=None, tol=None, tolNorm=None, n_iters=30, burn_in=30, maxiter=10, custom_convergence_f=None, iprint=False, full_output=False, learn_params_kwargs={'eta': 1, 'maxdlamda': 1}, generate_kwargs={})

Solve for maxent model parameters using MCH routine.

initial_guess : ndarray, None
Initial starting point.
constraints : ndarray, None
For debugging! Vector of correlations to fit.
tol : float, None
Maximum error allowed in any observable.
tolNorm : float, None
Norm error allowed in found solution.
n_iters : int, 30
Number of iterations to make between samples in MCMC sampling.
burn_in : int, 30
Initial burn in from random sample when MC sampling.
max_iter : int, 10
Max number of iterations of MC sampling and MCH approximation.
custom_convergence_f : function, None

Function for determining convergence criterion. At each iteration, this function should return the next set of learn_params_kwargs and optionally the sample size.

As an example: def learn_settings(i):

‘’’ Take in the iteration counter and set the maximum change allowed in any given parameter (maxdlamda) and the multiplicative factor eta, where d(parameter) = (error in observable) * eta.

Additional option is to also return the sample size for that step by returning a tuple. Larger sample sizes are necessary for higher accuracy. ‘’’ if i<10:

return {‘maxdlamda’:1,’eta’:1}
else:
return {‘maxdlamda’:.05,’eta’:.05}

iprint : bool, False full_output : bool, False

If True, also return the errflag and error history.

learn_parameters_kwargs : dict, {‘maxdlamda’:1,’eta’:1} generate_kwargs : dict, {}

ndarray
Solved multipliers (parameters). For Ising problem, these can be converted into matrix format using utils.vec2mat.
int
Error flag. 0, converged within given criterion 1, max iterations reached
ndarray
Log of errors in matching constraints at each step of iteration.
coniii.solvers.unwrap_self_worker_obj(arg, **kwarg)

coniii.samplers module

class coniii.samplers.HamiltonianMC(n, theta, calc_e, random_sample, grad_e=None, dt=0.01, leapfrogN=20, nCpus=0)

Bases: coniii.samplers.Sampler

generate_samples(nSamples, nBurn=100, fast=True, x0=None)

Generate nSamples from this Hamiltonian starting from random initial conditions from each sample.

sample(x0, nBurn, saveHistory=False)

Get a single sample by MC sampling from this Hamiltonian. Slow method

class coniii.samplers.Heisenberg3DSampler(J, calc_e, random_sample)

Bases: coniii.samplers.Sampler

Simple MC Sampling from Heisenberg model with a lot of helpful functions.

generate_samples() equilibrate_samples() sample_metropolis() sample_energy_min()

equilibrate_samples(samples, n_iters, method='mc', nCpus=0)
generate_samples(nSamples, n_iters=100, **kwargs)

sample_size : int

grad_E(X)

Gradient wrt theta and phi.

X : ndarray
with dims (nSpins,2) with angles theta and phi
sample_energy_min(nFixed=0, rng=RandomState(MT19937) at 0x7F74D4748160, initialState=None, method='powell', **kwargs)

Find local energy minimum given state in angular form. Angular representation makes it easy to be explicit about constraints on the vectors.

initialState : ndarray,None
n_samples x n_features x 2
nFixed : int,0
Number of vectors that are fixed.
sample_metropolis(oldState, E0)
s : ndarray
State to perturb randomly.
energy : float
Energy of configuration.
sample_nearby_sample(X, **kwargs)

Randomly move given state around for new metropolis sample. Question is whether it is more efficient to push only one of the many vectors around or all of them simultaneously.

sample_nearby_vector(v, nSamples=1, otheta=None, ophi=None, sigma=0.1)

Sample random vector that is nearby. It is important how you choose the width sigma. NOTE: code might be simplified by using arctan2 instead of arctan

v : ndarray
xyz vector about which to sample random vectors
nSamples : int,1
number of random samples
otheta : float,None
polar angle for v
ophi : float,None
azimuthal angle for v
sigma : float,.1
width of Gaussian about v
classmethod to_dict(data, names)

Convenience function taking 3d array of of samples and arranging them into n x 3 arrays in a dictionary.

class coniii.samplers.Metropolis(n, theta, calc_e=None, n_cpus=None, rng=None, boost=True, iprint=True)

Bases: coniii.samplers.Sampler

generate_cond_samples(sample_size, fixed_subset, burn_in=1000, n_cpus=None, initial_sample=None, systematic_iter=False, parallel=True)

Generate samples from conditional distribution (while a subset of the spins are held fixed). Samples are generated in parallel.

NOTE: There is a bug with multiprocess where many calls to the parallel sampling routine in a row leads to increasingly slow evaluation of the code.

sample_size : int fixed_subset : list of duples

Each duple is the index of the spin and the value to fix it at. These should be ordered by spin index.
burn_in : int
Burn in.
n_cpus : int
Number of cpus to use.
initial_sample : ndarray
Option to set initial random sample.
systematic_iter : bool
Iterate through spins systematically instead of choosing them randomly.
parallel : bool
If True, use parallelized routine.
ndarray
Samples from distribution.
ndarray
Energy of each sample.
generate_samples_boost(sample_size, n_iters=1000, burn_in=None, systematic_iter=False)

Generate Metropolis samples using C++ and boost.

sample_size : int
Number of samples.
n_iters : int, 1000
Number of Metropolis iterations between samples.
burn_in : int, None
If not set, will be the same value as n_iters.
systematic_iter : bool, False
If True, iterate through each element of system by increment index by one.
ndarray, optional
Saved array of energies at each sampling step.
generate_samples_parallel_boost(sample_size, n_iters=1000, burn_in=None, systematic_iter=False)

Generate samples in parallel. Each replica in self._samples runs on its own thread and a sample is generated every n_iters.

In order to control the random number generator, we pass in seeds that are samples from the class instance’s rng.

sample_size : int
Number of samples.
n_iters : int, 1000
Number of iterations between taking a random sample.
burn_in : int, None
If None, n_iters is used.
systematic_iter : bool, False
If True, iterate through spins systematically instead of choosing them randomly.
generate_samples_parallel_py(sample_size, n_iters=1000, burn_in=None, initial_sample=None, systematic_iter=False)

Generate samples in parallel. Each replica in self._samples runs on its own thread and a sample is generated every n_iters.

In order to control the random number generator, we pass in seeds that are samples from the class instance’s rng.

sample_size : int
Number of samples.
n_iters : int, 1000
Number of iterations between taking a random sample.
burn_in : int, None
If None, n_iters is used.
initial_sample : ndarray, None
Starting set of replicas otherwise self._samples is used.
systematic_iter : bool, False
If True, iterate through spins systematically instead of choosing them randomly.
generate_samples_py(sample_size, n_iters=1000, burn_in=None, systematic_iter=False, saveHistory=False, initial_sample=None)

Generate Metropolis samples using a for loop.

sample_size : int
Number of samples.
n_iters : int, 1000
Number of iterations to run the sampler floor.
burn_in : int, None
If None, n_iters is used.
systematic_iter : bool, False
If True, iterate through each element of system by increment index by one.
saveHistory : bool, False
If True, also save the energy of each sample at each sampling step.
initial_sample : ndarray, None
Start with this sample (i.e. to avoid warming up). Otherwise, self._samples is the initial sample.
ndarray, optional
Saved array of energies at each sampling step.
random_sample(n_samples)
sample_metropolis(sample0, E0, rng=None, flip_site=None, calc_e=None)

Metropolis sampling given an arbitrary energy function.

sample0 : ndarray
Sample to start with. Passed by ref and changed.
E0 : ndarray
Initial energy of state.
rng : np.random.RandomState
Random number generator.
flip_site : int
Site to flip.
calc_e : function
If another function to calculate energy should be used
float
delta energy.
update_parameters(theta)
class coniii.samplers.ParallelTempering(n, theta, calc_e, n_replicas, Tbds=(1.0, 3.0), sample_size=1000, replica_burnin=None, rep_ex_burnin=None, n_cpus=None, rng=None)

Bases: coniii.samplers.Sampler

burn_and_exchange(pool)

pool : mp.multiprocess.Pool

burn_in_replicas(pool=None, close_pool=True, n_iters=None)

Run each replica separately.

pool : multiprocess.Pool, None close_pool : bool, True

If True, call pool.close() at end.
n_iters : int, None
Default value is self.replicaBurnin.
generate_samples(sample_size, save_exchange_trajectory=False)

Burn in, run replica exchange simulation, then sample.

sample_size : int
Number of samples to take for each replica.
save_exchange_trajectory : bool, False
If True, keep track of the location of each replica in beta space and return the history.
ndarray, optional
Trajectory of each replica through beta space. Each row is tells where each index is located in beta space.
static initialize_beta(b0, b1, n_replicas)

Use linear interpolation of temperature range.

static iterate_beta(beta, acceptance_ratio)

Apply algorithm from Hukushima but reversed to maintain one replica at T=1.

beta : ndarray
Inverse temperature.
acceptance_ratio : ndarray
Estimate of acceptance ratio.
ndarray
New beta.
optimize_beta(n_samples, n_iters, tol=0.01, max_iter=10)

Find suitable temperature range for replicas. Sets self.beta.

n_samples : int
Number of samples to use to estimate acceptance ratio. Acceptance ratio is estimated as the average of these samples.
n_iters : int
Number of sampling iterations for each replica.
tol : float, .1
Average change in beta to reach before stopping.
max_iter : int, 10
Number of times to iterate algorithm for beta. Each iteration involves sampling from replicas.
setup_replicas()

Initialise a set of replicas at different temperatures using the Metropolis algorithm and optimize the temperatures. Replicas are burned in and ready to sample.

update_replica_parameters()

Update parameters for each replica. Remember that the parameters include the factor of beta.

class coniii.samplers.Potts3(n, theta, calc_e=None, n_cpus=None, rng=None, boost=True)

Bases: coniii.samplers.Metropolis

generate_samples_parallel_boost(sample_size, n_iters=1000, burn_in=None, systematic_iter=False)

Generate samples in parallel. Each replica in self._samples runs on its own thread and a sample is generated every n_iters.

In order to control the random number generator, we pass in seeds that are samples from the class instance’s rng.

sample_size : int
Number of samples.
n_iters : int, 1000
Number of iterations between taking a random sample.
burn_in : int, None
If None, n_iters is used.
systematic_iter : bool, False
If True, iterate through spins systematically instead of choosing them randomly.
random_sample(n_samples)
sample_metropolis(sample0, E0, rng=None, flip_site=None, calc_e=None)

Metropolis sampling given an arbitrary sampling function.

sample0 : ndarray
Sample to start with. Passed by ref and changed.
E0 : ndarray
Initial energy of state.
rng : np.random.RandomState
Random number generator.
flip_site : int
Site to flip.
calc_e : function
If another function to calculate energy should be used
float
delta energy.
class coniii.samplers.SWIsing(n, theta, calc_e, nCpus=None, rng=None)

Bases: coniii.samplers.Sampler

generate_sample(n_samples, n_iters, initial_state=None)

n_samples n_iters initial_state : ndarray,None

generate_sample_parallel(n_samples, n_iters, initial_state=None, n_cpus=None)

n_samples n_iters initial_state : ndarray,None

get_clusters(state)

Get a random sample of clusters.

one_step(state)
print_cluster_size(n_iters)
randomly_flip_clusters(state, clusters)
class coniii.samplers.Sampler(n, theta, **kwargs)

Bases: object

Base class for MCMC sampling.

generate_samples(sample_size, **kwargs)

sample_size : int

generate_samples_parallel(sample_size, **kwargs)

sample_size : int

sample_metropolis(s, energy)
s : ndarray
State to perturb randomly.
energy : float
Energy of configuration.
update_parameters(new_parameters)
class coniii.samplers.WolffIsing(J, h)

Bases: coniii.samplers.Sampler

build_cluster(state, initialsite)

Grow cluster from initial site.

find_neighbors(state, site, alreadyMarked)

Return neighbors of given site that need to be visited excluding sites that have already been visited. This is the implementation of the Wolff algorithm for finding neighbors such that detailed balance is satisfied. I have modified to include random fields such tha the probability of adding a neighbors depends both on its coupling with the current site and the neighbor’s magnetic field.

state site alreadyMarked

generate_sample(samplesize, n_iters, initialSample=None, save_history=False)

Generate samples by starting from random initial states.

generate_sample_parallel(samplesize, n_iters, initialSample=None)

Generate samples by starting from random or given initial states.

one_step(state, initialsite=None)

Run one iteration of the Wolff algorithm that involves finding a cluster and possibly flipping it.

update_parameters(J, h)
coniii.samplers.calc_e

Heisenberg model.

theta : ndarray
List of couplings Jij
x : ndarray
List of angles (theta_0,phi_0,theta_1,phi_1,…,theta_n,phi_n)
coniii.samplers.check_e_logp(sample, calc_e)

Boltzmann type model with discrete state space should have E propto -logP. Calculate these quantities for comparison.

sample calc_e

coniii.samplers.cross

Calculate the cross product of two 3d vectors.

coniii.samplers.cross_

Calculate the cross product of two 3d vectors.

coniii.samplers.grad_e

Derivatives wrt the angles of the spins.

coniii.samplers.grad_e_theta

Derivatives wrt the couplings theta.

coniii.samplers.iter_cluster(adj)

Cycle through all spins to get clusters.

coniii.samplers.iterate_neighbors

Iterate through all neighbors of a particular site and see if a bond should be formed between them.

n : int
System size.
ix : ndarray of bool
Indices of sites that have already been visited.
expdJ : ndarray
np.exp( -2*state[:,None]*state[None,:]*J )
r : ndarray
Array of random numbers.
coniii.samplers.jit_sample

Get a single sample by MC sampling from this Hamiltonian.

theta : ndarray
Parameters
x0 : ndarray
Sample

nBurn : int dt : float leapfrogN : int randNormal : ndarray

nBurn x ndim
randUnif : ndarray
nBurn
coniii.samplers.jit_sample_nearby_vector
coniii.samplers.pairwise_prod
coniii.samplers.sample_bonds
p : ndarray
Probability of bond formation.
r : ndarray
Random numbers.

state J

coniii.samplers.sample_ising(multipliers, n_samples, n_cpus=None, seed=None, generate_samples_kw={})

Easy way to Metropolis sample from Ising model.

multipliers : ndarray
N individual fields followed by N(N-1)/2 pairwise couplings.
n_samples : int
Number of samples to take.
n_cpus : int, None
Number of cpus to use. If more than one, multiprocessing invoked.
seed : int, None
Random number generator seed.
generate_samples_kw : dict, {}
Any extra arguments to send into Metropolis sampler. Default args are
n_iters=1000 systematic_iter=False saveHistory=False initial_sample=None
ndarray
Matrix of dimensions (n_samples, n).
coniii.samplers.spec_cluster(L, exact=True)
L : ndarray
Graph Laplacian

coniii.utils module

coniii.utils.adj

Return one-flip neighbors and a set of random neighbors. This is written to be used with the solvers.MPF class. Use adj_sym() if symmetric spins in {-1,1} are needed.

NOTE: For random neighbors, there is no check to make sure neighbors don’t repeat but this shouldn’t be a problem as long as state space is large enough.

s : ndarray
State whose neighbors are found. One-dimensional vector of spins.
n_random_neighbors : int,0
If >0, return this many random neighbors. Neighbors are just random states, but they are called “neighbors” because of the terminology in MPF. They can provide coupling from s to states that are very different, increasing the equilibration rate.
neighbors : ndarray
Each row is a neighbor. s.size + n_random_neighbors are returned.
coniii.utils.adj_sym

Symmetric version of adj() where spins are in {-1,1}.

coniii.utils.base_repr

Return decimal number in given base as list.

i : int base : int

list

coniii.utils.bin_states(n, sym=False)

Generate all possible binary spin states.

n : int
Number of spins.
sym : bool
If true, return states in {-1,1} basis.

v : ndarray

coniii.utils.calc_de(s, i)

Calculate the derivative of the energy wrt parameters given the state and index of the parameter. In this case, the parameters are the concatenated vector of {h_i,J_ij}.

s : ndarray
Two-dimensional vector of spins where each row is a state.

i : int

dE : float
Derivative of hamiltonian with respect to ith parameter, i.e. the corresponding observable.
coniii.utils.calc_overlap(sample, ignore_zeros=False)

<si_a si_b> between all pairs of replicas a and b

sample ignore_zeros (bool=False)

Instead of normalizing by the number of spins, normalize by the minimum number of nonzero spins.
coniii.utils.coarse_grain_with_func(X, n_times, sim_func, coarse_func)

Iteratively coarse-grain X by combining pairs with the highest similarity. Both the function to measure similarity and to implement the coarse-graining must be supplied.

X : ndarray
Each col is a variable and each row is an observation (n_samples, n_system).
n_times : int
Number of times to coarse grain.
sim_func : function
Takes an array like X and returns a vector of ncol*(ncol-1)//2 pairwise similarities.
coarse_func : function
Takes a two col array and returns a single vector.
ndarray
Coarse-grained version of X.
list of lists of ints
Each list specifies which columns of X have been coarse-grained into each col of the coarse X.
coniii.utils.convert_corr(si, sisj, convert_to, concat=False, **kwargs)

Convert single spin means and pairwise correlations between {0,1} and {-1,1} formulations.

si : ndarray
Individual means.
sisj : ndarray
Pairwise correlations.
convert_to : str
‘11’ will convert {0,1} formulation to +/-1 and ‘01’ will convert +/-1 formulation to {0,1}
concat : bool, False
If True, return concatenation of means and pairwise correlations.
ndarray
Averages <si>. Converted to appropriate basis. Returns concatenated vector <si> and <sisj> if concat is True.
ndarray, optional
Pairwise correlations <si*sj>. Converted to appropriate basis.
coniii.utils.convert_params(h, J, convert_to, concat=False)

Convert Ising model fields and couplings from {0,1} basis to {-1,1} and vice versa.

h : ndarray
Fields.
J : ndarray
Couplings.
convert_to : str
Either ‘01’ or ‘11’.
concat : bool, False
If True, return a vector concatenating fields and couplings.
ndarray
Mean bias h vector. Concatenated vector of h and J if concat is True.
ndarray, optional
Vector of J.
coniii.utils.define_ising_helper_functions()

Functions for plugging into solvers for +/-1 Ising model with fields h_i and couplings J_ij.

function
calc_e
function
calc_observables
function
mch_approximation
coniii.utils.define_ising_helper_functions_sym()

Functions for plugging into solvers for +/-1 Ising model with couplings J_ij and no fields.

function
calc_e
function
calc_observables
function
mch_approximation
coniii.utils.define_potts_helper_functions(k)

Helper functions for calculating quantities in k-state Potts model.

k : int
Number of possible states.
function
calc_e
function
calc_observables
function
mch_approximation
coniii.utils.define_pseudo_ising_helper_functions(N)

Define helper functions for using Pseudo method on Ising model.

N : int
System size.
function
get_multipliers_r
function
calc_observables_r
coniii.utils.define_pseudo_potts_helper_functions(n, k)

Define helper functions for using Pseudo method on Potts model with simple form for couplings that are only nonzero when the spins are occupying the same state.

n : int
System size.
k : int
Number of possible configurations in Potts model.
function
get_multipliers_r
function
calc_observables_r
coniii.utils.define_ternary_helper_functions()
coniii.utils.define_triplet_helper_functions()
coniii.utils.ind_to_sub

Convert index from flattened upper triangular matrix to pair subindex.

n : int
Dimension size of square array.
ix : int
Index to convert.
subix : tuple
(i,j)
coniii.utils.ising_convert_params(oparams, convert_to, concat=False)

General conversion of parameters from 01 to 11 basis.

Take set of Ising model parameters up to nth order interactions in either {0,1} or {-1,1} basis and convert to other basis.

oparams : tuple of lists
Tuple of lists of interactions between spins starting with the lowest order interactions. Each list should consist of all interactions of that order such that the length of each list should be binomial(n,i) for all i starting with i>=1.

convert_to : str concat : bool,False

params : tuple of lists or list
New parameters in order of lowest to highest order interactions to mean biases. Can all be concatenated together if concat switch is True.
coniii.utils.k_corr(X, k, weights=None, exclude_empty=False)

Calculate kth order correlations of spins.

X : ndarray
Dimensions (n_samples, n_dim).
k : int
Order of correlation function <s_{i_1} * s_{i_2} * … * s_{i_k}>.
weights : np.ndarray, None :
Calculate single and pairwise means given fractional weights for each state in the data such that a state only appears with some weight, typically less than one.
exclude_empty : bool, False
When using with {-1,1} basis, you can leave entries with 0 and those will not be counted for any pair. If True, the weights option doesn’t do anything.
ndarray
Kth order correlations <s_{i_1} * s_{i_2} * … * s_{i_k}>.
coniii.utils.mat2vec(multipliers)

Convert matrix form of Ising parameters to a vector.

This is specific to the Ising model.

multipliers : ndarray
Matrix of couplings with diagonal elements as fields.
ndarray
Vector of fields and couplings, respectively.
coniii.utils.multinomial(*args)
coniii.utils.pair_corr(X, weights=None, concat=False, exclude_empty=False, subtract_mean=False, laplace_count=False)

Calculate averages and pairwise correlations of spins.

X : ndarray
Dimensions (n_samples,n_dim).
weights : float or np.ndarray or twople, None
If an array is passed, it must be the length of the data and each data point will be given the corresponding weight. Otherwise, the two element tuple should contain the normalization for each mean and each pairwise correlation, in that order. In other words, the first array should be length {s_i} and the second length {si*s_j}.
concat : bool, False
Return means concatenated with the pairwise correlations into one array.
exclude_empty : bool, False
When using with {-1,1} basis, you can leave entries with 0 and those will not be counted for any pair. If True, the weights option doesn’t do anything.
subtract_mean : bool, False
If True, return pairwise correlations with product of individual means subtracted.

laplace_count :

twople
(si,sisj) or np.concatenate((si,sisj))
coniii.utils.replace_diag(mat, newdiag)

Replace diagonal entries of square matrix.

mat : ndarray newdiag : ndarray

ndarray

coniii.utils.split_concat_params(p, n)

Split parameters for Ising model that have all been concatenated together into a single list into separate lists. Assumes that the parameters are increasing in order of interaction and that all parameters are present.

p : list-like

list of list-like
Parameters increasing in order: (h, Jij, Kijk, … ).
coniii.utils.state_probs(v, allstates=None, weights=None, normalized=True)

Get probability of unique states. There is an option to allow for weighted counting.

states : ndarray
Sample of states on which to extract probabilities of unique configurations with dimensions (n_samples,n_dimension).
allstates : ndarray, None
Unique configurations to look for with dimensions (n_samples, n_dimension).
weights : vector, None
For weighted counting of each state given in allstate kwarg.
normalized : bool, True
If True, return probability distribution instead of frequency count.
ndarray
Vector of the probabilities of each state.
ndarray
All unique states found in the data. Each state is a row. Only returned if allstates kwarg is not provided.
coniii.utils.sub_to_ind

Convert pair of coordinates of a symmetric square array into consecutive index of flattened upper triangle. This is slimmed down so it won’t throw errors like if i>n or j>n or if they’re negative. Only checking for if the returned index is negative which could be problematic with wrapped indices.

n : int
Dimension of square array
i,j : int
coordinates

int

coniii.utils.unique_rows(mat, return_inverse=False)

Return unique rows indices of a numeric numpy array.

mat : ndarray return_inverse : bool

If True, return inverse that returns back indices of unique array that would return the original array
u : ndarray
Unique elements of matrix.
idx : ndarray
row indices of given mat that will give unique array
coniii.utils.unravel_index(ijk, n)

Unravel multi-dimensional index to flattened index but specifically for multi-dimensional analog of an upper triangular array (lower triangle indices are not counted).

ijk : tuple
Raveled index to unravel.
n : int
System size.
ix : int
Unraveled index.
coniii.utils.vec2mat(multipliers, separate_fields=False)

Convert vector of parameters containing fields and couplings to a matrix where the diagonal elements are the fields and the remaining elements are the couplings. Fields can be returned separately with the separate_fields keyword argument.

This is specific to the Ising model.

multipliers : ndarray
Vector of fields and couplings.

separate_fields : bool, False

ndarray
n x n matrix. Diagonal elements are fields unless separate_fields keyword argument is True, in which case the diagonal elements are 0.
ndarray (optional)
Fields if separate_fields keyword argument is True.
coniii.utils.xbin_states(n, sym=False)

Generator for iterating through all possible binary states.

n : int
Number of spins.
sym : bool
If true, return states in {-1,1} basis.

generator

coniii.utils.xpotts_states

Generator for iterating through all states for Potts model with k distinct states. This is a faster version of calling xbin_states(n, False) except with strings returned as elements instead of integers.

n : int
Number of spins.
k : int
Number of distinct states. These are labeled by integers starting from 0 and must be <=36.

generator

coniii.utils.zero_diag(mat)

Replace diagonal entries of square matrix with zeros.

mat : ndarray

ndarray

Indices and tables