# coding: utf-8
from __future__ import division, print_function
__author__ = "adrn <adrn@astro.columbia.edu>"
# Third-party
import astropy.units as u
import numpy as np
from scipy.signal import argrelmin
# Project
from . import CartesianPhaseSpacePosition, CartesianOrbit
__all__ = ['fast_lyapunov_max', 'lyapunov_max', 'surface_of_section']
[docs]def fast_lyapunov_max(w0, potential, dt, n_steps, d0=1e-5,
n_steps_per_pullback=10, noffset_orbits=2, t1=0.,
atol=1E-10, rtol=1E-10, nmax=0, return_orbit=True):
"""
Compute the maximum Lyapunov exponent using a C-implemented estimator
that uses the DOPRI853 integrator.
Parameters
----------
w0 : `~gala.dynamics.PhaseSpacePosition`, array_like
Initial conditions.
dt : numeric
Timestep.
n_steps : int
Number of steps to run for.
d0 : numeric (optional)
The initial separation.
n_steps_per_pullback : int (optional)
Number of steps to run before re-normalizing the offset vectors.
noffset_orbits : int (optional)
Number of offset orbits to run.
t1 : numeric (optional)
Time of initial conditions. Assumed to be t=0.
return_orbit : bool (optional)
Store the full orbit for the parent and all offset orbits.
Returns
-------
LEs : :class:`~astropy.units.Quantity`
Lyapunov exponents calculated from each offset / deviation orbit.
orbit : `~gala.dynamics.CartesianOrbit` (optional)
"""
from .lyapunov import dop853_lyapunov_max, dop853_lyapunov_max_dont_save
if not hasattr(potential, 'c_instance'):
raise TypeError("Input potential must be a CPotential subclass.")
if not isinstance(w0, CartesianPhaseSpacePosition):
w0 = np.asarray(w0)
ndim = w0.shape[0]//2
w0 = CartesianPhaseSpacePosition(pos=w0[:ndim],
vel=w0[ndim:])
_w0 = np.squeeze(w0.w(potential.units))
if _w0.ndim > 1:
raise ValueError("Can only compute fast Lyapunov exponent for a single orbit.")
if return_orbit:
t,w,l = dop853_lyapunov_max(potential.c_instance, _w0,
dt, n_steps+1, t1,
d0, n_steps_per_pullback, noffset_orbits,
atol, rtol, nmax)
w = np.rollaxis(w, -1)
try:
tunit = potential.units['time']
except (TypeError, AttributeError):
tunit = u.dimensionless_unscaled
orbit = CartesianOrbit.from_w(w=w, units=potential.units, t=t*tunit, potential=potential)
return l/tunit, orbit
else:
l = dop853_lyapunov_max_dont_save(potential.c_instance, _w0,
dt, n_steps+1, t1,
d0, n_steps_per_pullback, noffset_orbits,
atol, rtol, nmax)
try:
tunit = potential.units['time']
except (TypeError, AttributeError):
tunit = u.dimensionless_unscaled
return l/tunit
[docs]def lyapunov_max(w0, integrator, dt, n_steps, d0=1e-5, n_steps_per_pullback=10,
noffset_orbits=8, t1=0., units=None):
"""
Compute the maximum Lyapunov exponent of an orbit by integrating many
nearby orbits (``noffset``) separated with isotropically distributed
directions but the same initial deviation length, ``d0``. This algorithm
re-normalizes the offset orbits every ``n_steps_per_pullback`` steps.
Parameters
----------
w0 : `~gala.dynamics.PhaseSpacePosition`, array_like
Initial conditions.
integrator : `~gala.integrate.Integrator`
An instantiated `~gala.integrate.Integrator` object. Must have a run() method.
dt : numeric
Timestep.
n_steps : int
Number of steps to run for.
d0 : numeric (optional)
The initial separation.
n_steps_per_pullback : int (optional)
Number of steps to run before re-normalizing the offset vectors.
noffset_orbits : int (optional)
Number of offset orbits to run.
t1 : numeric (optional)
Time of initial conditions. Assumed to be t=0.
units : `~gala.units.UnitSystem` (optional)
If passing in an array (not a `~gala.dynamics.PhaseSpacePosition`),
you must specify a unit system.
Returns
-------
LEs : :class:`~astropy.units.Quantity`
Lyapunov exponents calculated from each offset / deviation orbit.
orbit : `~gala.dynamics.CartesianOrbit`
"""
if units is not None:
pos_unit = units['length']
vel_unit = units['length']/units['time']
else:
pos_unit = u.dimensionless_unscaled
vel_unit = u.dimensionless_unscaled
if not isinstance(w0, CartesianPhaseSpacePosition):
w0 = np.asarray(w0)
ndim = w0.shape[0]//2
w0 = CartesianPhaseSpacePosition(pos=w0[:ndim]*pos_unit,
vel=w0[ndim:]*vel_unit)
_w0 = w0.w(units)
ndim = 2*w0.ndim
# number of iterations
niter = n_steps // n_steps_per_pullback
# define offset vectors to start the offset orbits on
d0_vec = np.random.uniform(size=(ndim,noffset_orbits))
d0_vec /= np.linalg.norm(d0_vec, axis=0)[np.newaxis]
d0_vec *= d0
w_offset = _w0 + d0_vec
all_w0 = np.hstack((_w0,w_offset))
# array to store the full, main orbit
full_w = np.zeros((ndim,n_steps+1,noffset_orbits+1))
full_w[:,0] = all_w0
full_ts = np.zeros((n_steps+1,))
full_ts[0] = t1
# arrays to store the Lyapunov exponents and times
LEs = np.zeros((niter,noffset_orbits))
ts = np.zeros_like(LEs)
time = t1
for i in range(1,niter+1):
ii = i * n_steps_per_pullback
orbit = integrator.run(all_w0, dt=dt, n_steps=n_steps_per_pullback, t1=time)
tt = orbit.t.value
ww = orbit.w(units)
time += dt*n_steps_per_pullback
main_w = ww[:,-1,0:1]
d1 = ww[:,-1,1:] - main_w
d1_mag = np.linalg.norm(d1, axis=0)
LEs[i-1] = np.log(d1_mag/d0)
ts[i-1] = time
w_offset = ww[:,-1,0:1] + d0 * d1 / d1_mag[np.newaxis]
all_w0 = np.hstack((ww[:,-1,0:1],w_offset))
full_w[:,(i-1)*n_steps_per_pullback+1:ii+1] = ww[:,1:]
full_ts[(i-1)*n_steps_per_pullback+1:ii+1] = tt[1:]
LEs = np.array([LEs[:ii].sum(axis=0)/ts[ii-1] for ii in range(1,niter)])
try:
t_unit = units['time']
except (TypeError, AttributeError):
t_unit = u.dimensionless_unscaled
orbit = CartesianOrbit.from_w(w=full_w, units=units, t=full_ts*t_unit)
return LEs/t_unit, orbit
[docs]def surface_of_section(orbit, plane_ix, interpolate=False):
"""
Generate and return a surface of section from the given orbit.
.. warning::
This is an experimental function and the API may change.
Parameters
----------
orbit : `~gala.dynamics.CartesianOrbit`
plane_ix : int
Integer that represents the coordinate to record crossings in. For
example, for a 2D Hamiltonian where you want to make a SoS in
:math:`y-p_y`, you would specify ``plane_ix=0`` (crossing the
:math:`x` axis), and this will only record crossings for which
:math:`p_x>0`.
interpolate : bool (optional)
Whether or not to interpolate on to the plane of interest. This
makes it much slower, but will work for orbits with a coarser
sampling.
Returns
-------
Examples
--------
If your orbit of interest is a tube orbit, it probably conserves (at
least approximately) some equivalent to angular momentum in the direction
of the circulation axis. Therefore, a surface of section in R-z should
be instructive for classifying these orbits. TODO...show how to convert
an orbit to Cylindrical..etc...
"""
w = orbit.w()
if w.ndim == 2:
w = w[...,None]
ndim,ntimes,norbits = w.shape
H_dim = ndim // 2
p_ix = plane_ix + H_dim
if interpolate:
raise NotImplementedError("Not yet implemented, sorry!")
# record position on specified plane when orbit crosses
all_sos = np.zeros((ndim,norbits), dtype=object)
for n in range(norbits):
cross_ix = argrelmin(w[plane_ix,:,n]**2)[0]
cross_ix = cross_ix[w[p_ix,cross_ix,n] > 0.]
sos = w[:,cross_ix,n]
for j in range(ndim):
all_sos[j,n] = sos[j,:]
return all_sos