def _reconstruct(self, observation, out):
observation = self.observation_space.element(observation)
out[:] = self.x0
if self.op_is_symmetric:
iterative.conjugate_gradient(self.op, out, observation,
self.niter, self.callback)
else:
iterative.conjugate_gradient_normal(self.op, out, observation,
self.niter, self.callback)
return out
def _reconstruct(self, observation, out):
observation = self.observation_space.element(observation)
out_ = out
if out not in self.reco_space:
out_ = self.reco_space.zero()
out_[:] = self.x0
if self.op_is_symmetric:
iterative.conjugate_gradient(self.op, out_, observation,
self.niter, self.callback)
else:
iterative.conjugate_gradient_normal(self.op, out_, observation,
self.niter, self.callback)
if out not in self.reco_space:
out[:] = out_
return out
def newtons_method(op, x, line_search, num_iter=10, cg_iter=None,
partial=None):
"""Newton's method for solving a system of equations.
This is a general and optimized implementation of Newton's method
for solving the problem::
f(x) = 0
of finding a root of a function.
The algorithm is well-known and there is a vast literature about it.
Among others, the method is described in [BV2004]_, Sections 9.5
and 10.2 (`book available online
<http://stanford.edu/~boyd/cvxbook/bv_cvxbook.pdf>`_),
[GNS2009]_, Section 2.7 for solving nonlinear equations and Section
11.3 for its use in minimization, and wikipedia on `Newton's_method
<https://en.wikipedia.org/wiki/Newton's_method>`_.
Parameters
----------
op : `Operator`
Gradient of the objective function, ``x --> grad f(x)``
x : element in the domain of ``op``
Starting point of the iteration
line_search : `LineSearch`
Strategy to choose the step length
num_iter : `int`, optional
Number of iterations
cg_iter : `int`, optional
Number of iterations in the the conjugate gradient solver,
for computing the search direction.
partial : `Partial`, optional
Object executing code per iteration, e.g. plotting each iterate
Notes
----------
The algorithm works by iteratively solving
:math:`\partial f(x_k)p_k = -f(x_k)`
and then updating as
:math:`x_{k+1} = x_k + \\alpha x_k`,
where :math:`\\alpha` is a suitable step length (see the
references). In this implementation the system of equations are
solved using the conjugate gradient method.
"""
# TODO: update doc
if cg_iter is None:
# Motivated by that if it is Ax = b, x and b in Rn, it takes at most n
# iterations to solve with cg
cg_iter = op.domain.size
# TODO: optimize by using lincomb and avoiding to create copies
for _ in range(num_iter):
# Initialize the search direction to 0
search_direction = x.space.zero()
# Compute hessian (as operator) and gradient in the current point
hessian = op.derivative(x)
deriv_in_point = op(x).copy()
# Solving A*x = b for x, in this case f'(x)*p = -f(x)
# TODO: Let the user provide/choose method for how to solve this?
conjugate_gradient(hessian, search_direction,
-1 * deriv_in_point, cg_iter)
# Computing step length
dir_deriv = search_direction.inner(deriv_in_point)
step_length = line_search(x, search_direction, dir_deriv)
# Updating
x += step_length * search_direction
if partial is not None:
partial(x)
def newtons_method(f, x, line_search=1.0, maxiter=1000, tol=1e-16,
cg_iter=None, callback=None):
"""Newton's method for minimizing a functional.
Notes
-----
This is a general and optimized implementation of Newton's method
for solving the problem:
.. math::
\min f(x)
for a differentiable function
:math:`f: \mathcal{X}\\to \mathbb{R}` on a Hilbert space
:math:`\mathcal{X}`. It does so by finding a zero of the gradient
.. math::
\\nabla f: \mathcal{X} \\to \mathcal{X}.
of finding a root of a function.
The algorithm is well-known and there is a vast literature about it.
Among others, the method is described in [BV2004], Sections 9.5
and 10.2 (`book available online
<http://stanford.edu/~boyd/cvxbook/bv_cvxbook.pdf>`_),
[GNS2009], Section 2.7 for solving nonlinear equations and Section
11.3 for its use in minimization, and wikipedia on `Newton's_method
<https://en.wikipedia.org/wiki/Newton's_method>`_.
The algorithm works by iteratively solving
.. math::
\partial f(x_k)p_k = -f(x_k)
and then updating as
.. math::
x_{k+1} = x_k + \\alpha x_k,
where :math:`\\alpha` is a suitable step length (see the
references). In this implementation the system of equations are
solved using the conjugate gradient method.
Parameters
----------
f : `Functional`
Goal functional. Needs to have ``f.gradient`` and
``f.gradient.derivative``.
x : ``op.domain`` element
Starting point of the iteration
line_search : float or `LineSearch`, optional
Strategy to choose the step length. If a float is given, uses it as a
fixed step length.
maxiter : int, optional
Maximum number of iterations.
tol : float, optional
Tolerance that should be used for terminating the iteration.
cg_iter : int, optional
Number of iterations in the the conjugate gradient solver,
for computing the search direction.
callback : callable, optional
Object executing code per iteration, e.g. plotting each iterate
References
----------
[BV2004] Boyd, S, and Vandenberghe, L. *Convex optimization*.
Cambridge university press, 2004.
[GNS2009] Griva, I, Nash, S G, and Sofer, A. *Linear and nonlinear
optimization*. Siam, 2009.
"""
# TODO: update doc
grad = f.gradient
if x not in grad.domain:
raise TypeError('`x` {!r} is not in the domain of `f` {!r}'
''.format(x, grad.domain))
if not callable(line_search):
line_search = ConstantLineSearch(line_search)
if cg_iter is None:
# Motivated by that if it is Ax = b, x and b in Rn, it takes at most n
# iterations to solve with cg
cg_iter = grad.domain.size
# TODO: optimize by using lincomb and avoiding to create copies
for _ in range(maxiter):
# Initialize the search direction to 0
search_direction = x.space.zero()
# Compute hessian (as operator) and gradient in the current point
hessian = grad.derivative(x)
deriv_in_point = grad(x)
# Solving A*x = b for x, in this case f''(x)*p = -f'(x)
# TODO: Let the user provide/choose method for how to solve this?
try:
hessian_inverse = hessian.inverse
except NotImplementedError:
conjugate_gradient(hessian, search_direction,
-deriv_in_point, cg_iter)
else:
hessian_inverse(-deriv_in_point, out=search_direction)
# Computing step length
dir_deriv = search_direction.inner(deriv_in_point)
if np.abs(dir_deriv) <= tol:
return
step_length = line_search(x, search_direction, dir_deriv)
# Updating
x += step_length * search_direction
if callback is not None:
callback(x)
def newtons_method(f, x, line_search=1.0, maxiter=1000, tol=1e-16,
cg_iter=None, callback=None):
"""Newton's method for minimizing a functional.
Notes
-----
This is a general and optimized implementation of Newton's method
for solving the problem:
:math:`\min f(x)`
for a differentiable function
:math:`f: \mathcal{X}\\to \mathbb{R}` on a Hilbert space
:math:`\mathcal{X}`. It does so by finding a zero of the gradient
:math:`\\nabla f: \mathcal{X} \\to \mathcal{X}`.
of finding a root of a function.
The algorithm is well-known and there is a vast literature about it.
Among others, the method is described in [BV2004]_, Sections 9.5
and 10.2 (`book available online
<http://stanford.edu/~boyd/cvxbook/bv_cvxbook.pdf>`_),
[GNS2009]_, Section 2.7 for solving nonlinear equations and Section
11.3 for its use in minimization, and wikipedia on `Newton's_method
<https://en.wikipedia.org/wiki/Newton's_method>`_.
The algorithm works by iteratively solving
:math:`\partial f(x_k)p_k = -f(x_k)`
and then updating as
:math:`x_{k+1} = x_k + \\alpha x_k`,
where :math:`\\alpha` is a suitable step length (see the
references). In this implementation the system of equations are
solved using the conjugate gradient method.
Parameters
----------
f : `Functional`
Goal functional. Needs to have ``f.gradient`` and
``f.gradient.derivative``.
x : ``op.domain`` element
Starting point of the iteration
line_search : float or `LineSearch`, optional
Strategy to choose the step length. If a float is given, uses it as a
fixed step length.
maxiter : int, optional
Maximum number of iterations.
tol : float, optional
Tolerance that should be used for terminating the iteration.
cg_iter : int, optional
Number of iterations in the the conjugate gradient solver,
for computing the search direction.
callback : callable, optional
Object executing code per iteration, e.g. plotting each iterate
"""
# TODO: update doc
grad = f.gradient
if x not in grad.domain:
raise TypeError('`x` {!r} is not in the domain of `f` {!r}'
''.format(x, grad.domain))
if not callable(line_search):
line_search = ConstantLineSearch(line_search)
if cg_iter is None:
# Motivated by that if it is Ax = b, x and b in Rn, it takes at most n
# iterations to solve with cg
cg_iter = grad.domain.size
# TODO: optimize by using lincomb and avoiding to create copies
for _ in range(maxiter):
# Initialize the search direction to 0
search_direction = x.space.zero()
# Compute hessian (as operator) and gradient in the current point
hessian = grad.derivative(x)
deriv_in_point = grad(x)
# Solving A*x = b for x, in this case f''(x)*p = -f'(x)
# TODO: Let the user provide/choose method for how to solve this?
conjugate_gradient(hessian, search_direction,
-deriv_in_point, cg_iter)
# Computing step length
dir_deriv = search_direction.inner(deriv_in_point)
if np.abs(dir_deriv) <= tol:
return
step_length = line_search(x, search_direction, dir_deriv)
# Updating
x += step_length * search_direction
if callback is not None:
callback(x)
def newtons_method(op,
x,
line_search,
num_iter=10,
cg_iter=None,
callback=None):
"""Newton's method for solving a system of equations.
This is a general and optimized implementation of Newton's method
for solving the problem::
f(x) = 0
of finding a root of a function.
The algorithm is well-known and there is a vast literature about it.
Among others, the method is described in [BV2004]_, Sections 9.5
and 10.2 (`book available online
<http://stanford.edu/~boyd/cvxbook/bv_cvxbook.pdf>`_),
[GNS2009]_, Section 2.7 for solving nonlinear equations and Section
11.3 for its use in minimization, and wikipedia on `Newton's_method
<https://en.wikipedia.org/wiki/Newton's_method>`_.
Parameters
----------
op : `Operator`
Gradient of the objective function, ``x --> grad f(x)``
x : element in the domain of ``op``
Starting point of the iteration
line_search : `LineSearch`
Strategy to choose the step length
num_iter : `int`, optional
Number of iterations
cg_iter : `int`, optional
Number of iterations in the the conjugate gradient solver,
for computing the search direction.
callback : `callable`, optional
Object executing code per iteration, e.g. plotting each iterate
Notes
----------
The algorithm works by iteratively solving
:math:`\partial f(x_k)p_k = -f(x_k)`
and then updating as
:math:`x_{k+1} = x_k + \\alpha x_k`,
where :math:`\\alpha` is a suitable step length (see the
references). In this implementation the system of equations are
solved using the conjugate gradient method.
"""
# TODO: update doc
if cg_iter is None:
# Motivated by that if it is Ax = b, x and b in Rn, it takes at most n
# iterations to solve with cg
cg_iter = op.domain.size
# TODO: optimize by using lincomb and avoiding to create copies
for _ in range(num_iter):
# Initialize the search direction to 0
search_direction = x.space.zero()
# Compute hessian (as operator) and gradient in the current point
hessian = op.derivative(x)
deriv_in_point = op(x).copy()
# Solving A*x = b for x, in this case f'(x)*p = -f(x)
# TODO: Let the user provide/choose method for how to solve this?
conjugate_gradient(hessian, search_direction, -1 * deriv_in_point,
cg_iter)
# Computing step length
dir_deriv = search_direction.inner(deriv_in_point)
step_length = line_search(x, search_direction, dir_deriv)
# Updating
x += step_length * search_direction
if callback is not None:
callback(x)
