affine.factored_matrix¶
Module: affine.factored_matrix¶
Inheritance diagram for regreg.affine.factored_matrix:
This module contains the implementation operator and nuclear norms, used in matrix completion problems and other low-rank factorization problems.
Classes¶
factored_matrix¶
-
class
regreg.affine.factored_matrix.factored_matrix(linear_operator, min_singular=0.0, tol=1e-05, initial_rank=None, initial=None, affine_offset=None, debug=False)¶ Bases:
objectA class for storing the SVD of a linear_tranform. Has affine_transform attributes like linear_map.
-
__init__(linear_operator, min_singular=0.0, tol=1e-05, initial_rank=None, initial=None, affine_offset=None, debug=False)¶ Initialize self. See help(type(self)) for accurate signature.
-
property
X¶
-
adjoint_map(x)¶
-
affine_map(x)¶
-
copy()¶
-
property
factors¶
-
linear_map(x)¶
-
offset_map(x, copy=False)¶
-
nuclear_norm¶
-
class
regreg.affine.factored_matrix.nuclear_norm(shape, lagrange=None, bound=None, offset=None, quadratic=None, initial=None, initial_rank=10, warm_start=None)¶ Bases:
regreg.atoms.svd_norms.nuclear_normThe nuclear norm
-
__init__(shape, lagrange=None, bound=None, offset=None, quadratic=None, initial=None, initial_rank=10, warm_start=None)¶ Initialize self. See help(type(self)) for accurate signature.
-
classmethod
affine(linear_operator, offset, lagrange=None, diag=False, bound=None, quadratic=None)¶ This is the same as the linear class method but with offset as a positional argument
-
apply_offset(x)¶ If self.offset is not None, return x-self.offset, else return x.
-
property
bound¶
-
bound_prox(X, bound=None)¶ Return unique minimizer
\[{B}^{\delta}(\theta) = \text{argmin}_{B \in \mathbb{R}^{{n \times p}}} \frac{1}{2} \|\theta-B\|^2_2 \ \text{s.t.} \ \|B\|_* \leq \delta\]where \(\delta\) is the bound parameter and \(\theta\) is arg.
If the argument bound is None and the atom is in bound mode,
self.boundis used as the bound parameter, else an exception is raised.The class atom’s bound_prox just returns the appropriate bound parameter for use by the subclasses.
- Parameters
arg : np.ndarray(np.float)
Argument of the proximal map.
bound : float (optional)
Bound for the constraint on the seminorm. Defaults to self.bound.
- Returns
Z : np.ndarray(np.float)
The proximal map of arg.
-
static
check_subgradient(atom, prox_center)¶ For a given seminorm, verify the KKT condition for the problem for the proximal problem
\[\text{minimize}_u \frac{1}{2} \|u-z\|^2_2 + h(z)\]where \(z\) is the prox_center and \(h\) is atom which may be in Lagrange or bound form.
If the atom is in Lagrange form, this function should return two values equal to the seminorm of the minimizer. If it is bound form it should return two values equal to the dual seminorm of the residual, i.e. the prox_center minus the minimizer.
- Parameters
atom : seminorm
prox_center : np.ndarray(np.float)
Center for the proximal map.
- Returns
v1, v2 : float
Two values that should be equal if the proximal map is correct.
-
property
conjugate¶ The conjugate of an atom.
-
constraint(X, bound=None)¶ Verify \(\|B\|_* \leq \delta\), where \(\delta\) is bound.
If the result is True, returns 0, else returns np.inf.
The class seminorm’s constraint just returns the appropriate bound parameter for use by the subclasses.
-
property
dual¶
-
get_bound()¶ Get method of the bound property.
>>> import regreg.api as rr >>> constraint = rr.nuclear_norm((5,4), bound=2.3) >>> constraint.bound 2.3
-
get_conjugate()¶ Return the conjugate of an given atom.
>>> import regreg.api as rr >>> penalty = rr.nuclear_norm((5,4), lagrange=3.4) >>> penalty.get_conjugate() operator_norm(..., bound=3.4...)
-
get_dual()¶ Return the dual of an atom. This dual is formed by making introducing new variables \(v=Ax\) where \(A\) is self.linear_transform.
>>> import regreg.api as rr >>> penalty = rr.nuclear_norm((5,4), lagrange=2.3) >>> penalty nuclear_norm(..., lagrange=2.3...) >>> penalty.dual (<regreg.affine.identity object at 0x...>, operator_norm(..., bound=2.3...))
If there is a linear part to the penalty, the linear_transform may not be identity. For example, the 1D fused LASSO penalty:
>>> D = (np.identity(4) + np.diag(-np.ones(3),1))[:-1] >>> D array([[ 1., -1., 0., 0.], [ 0., 1., -1., 0.], [ 0., 0., 1., -1.]]) >>> linear_atom = rr.l1norm.linear(D, lagrange=2.3) >>> linear_atom affine_atom(l1norm((3,), lagrange=2.3...), array([[ 1., -1., 0., 0.], [ 0., 1., -1., 0.], [ 0., 0., 1., -1.]])) >>> linear_atom.dual (<regreg.affine.linear_transform object at 0x...>, supnorm((3,), bound=2.3...))
-
get_lagrange()¶ Get method of the lagrange property.
>>> import regreg.api as rr >>> penalty = rr.nuclear_norm((5,4), lagrange=3.4) >>> penalty.lagrange 3.4
-
get_offset()¶
-
get_quadratic()¶ Get the quadratic part of the composite.
-
property
lagrange¶
-
lagrange_prox(X, lipschitz=1, lagrange=None)¶ Return unique minimizer
\[{B}^{\lambda}(\theta) = \text{argmin}_{B \in \mathbb{R}^{{n \times p}}} \frac{L}{2} \|\theta-B\|^2_2 + \lambda \|B\|_* \]Above, \(\lambda\) is the Lagrange parameter and \(L\) is the Lipschitz parameter and \(\theta\) is arg.
If the argument lagrange is None and the atom is in lagrange mode,
self.lagrangeis used as the lagrange parameter, else an exception is raised.The class atom’s lagrange_prox just returns the appropriate lagrange parameter for use by the subclasses.
- Parameters
arg : np.ndarray(np.float)
Argument of the proximal map.
lipschitz : float
Coefficient in front of the quadratic.
lagrange : float (optional)
Lagrange factor in front of the seminorm. Defaults to self.lagrange.
- Returns
Z : np.ndarray(np.float)
The proximal map of arg.
-
latexify(var=None, idx='')¶ Return a LaTeX representation of an object.
>>> import regreg.api as rr >>> penalty = rr.l1norm(10, lagrange=0.9) >>> penalty.latexify(var=r'\gamma') '\\lambda_{} \\|\\gamma\\|_1'
- Parameters
var : string
Argument of the functions
idx : string
Optional subscript index.
- Returns
L : string
A LaTeX representation of the atom.
-
classmethod
linear(linear_operator, lagrange=None, diag=False, bound=None, quadratic=None, offset=None)¶
-
property
linear_transform¶ The linear transform applied before a penalty is computed. Defaults to regreg.affine.identity
>>> from regreg.api import l1norm >>> penalty = l1norm(30, lagrange=3.4) >>> type(penalty.linear_transform) <class 'regreg.affine.identity'>
-
nonsmooth_objective(arg, check_feasibility=False)¶ The nonsmooth objective function of the atom. Includes self.quadratic.objective(arg).
>>> import regreg.api as rr >>> penalty = rr.l1norm(4, lagrange=2) >>> penalty.nonsmooth_objective([3, 4, 5, 9]) 42.0 >>> 2 * sum([3, 4, 5, 9]) 42
- Parameters
arg : np.ndarray(np.float)
Argument of the seminorm.
check_feasibility : bool
If True, then return np.inf if appropriate.
- Returns
value : np.float
The seminorm of arg.
-
objective(x, check_feasibility=False)¶
-
objective_template= '\\|%(var)s\\|_*'¶
-
objective_vars= {'dualnormklass': 'operator_norm', 'initargs': '(5,4)', 'linear': 'D', 'normklass': 'nuclear_norm', 'offset': '\\alpha', 'shape': '{n \\times p}', 'var': 'B'}¶
-
property
offset¶
-
prox_tol= 1e-05¶
-
proximal(quadratic, prox_control=None)¶ The proximal operator. If the atom is in Lagrange mode, this has the form
\[B^{\lambda}(\theta) = \text{argmin}_{B \in \mathbb{R}^{{n \times p}}} \frac{L}{2} \|\theta-B\|^2_2 + \lambda h(B-\alpha) + \langle B, \eta \rangle\]where \(\alpha\) is self.offset, \(\eta\) is quadratic.linear_term, \(\theta\) is quadratic.center and
\[h(B) = \|B\|_*\]If the atom is in bound mode, then this has the form
\[B^{\delta}(\theta) = \text{argmin}_{B \in \mathbb{R}^{{n \times p}}} \frac{L}{2} \|\theta-B\|^2_2 + \langle B, \eta \rangle \ \text{s.t.} \ h(B - \alpha) \leq \delta\]>>> import regreg.api as rr >>> penalty = rr.l1norm(4, lagrange=2) >>> Q = rr.identity_quadratic(1.5, [3, -4, -1, 1], 0, 0) >>> penalty.proximal(Q) array([ 1.6666..., -2.6666..., -0. , 0. ])
- Parameters
quadratic : regreg.identity_quadratic.identity_quadratic
A quadratic added to the atom before minimizing.
prox_control : [None, dict]
This argument is ignored for seminorms, but otherwise is passed to regreg.algorithms.FISTA if the atom needs to be solved iteratively.
- Returns
Z : np.ndarray(np.float)
The proximal map of the implied center of quadratic.
-
proximal_optimum(quadratic)¶
-
proximal_step(quadratic, prox_control=None)¶ Compute the proximal optimization
- Parameters
prox_control: [None, dict]
If not None, then a dictionary of parameters for the prox procedure
-
property
quadratic¶ Quadratic part of the object, instance of regreg.identity_quadratic.identity_quadratic.
-
seminorm(X, check_feasibility=False, lagrange=None)¶ Return \(\lambda \cdot \|B\|_*\), where \(\lambda\) is lagrange. If check_feasibility is True, and seminorm is unbounded, will return np.inf if appropriate.
The class seminorm’s seminorm just returns the appropriate lagrange parameter for use by the subclasses.
-
set_bound(bound)¶ Set method of the bound property.
>>> import regreg.api as rr >>> constraint = rr.nuclear_norm((5,4), bound=3.4) >>> constraint.bound 3.4 >>> constraint.bound = 2.3 >>> constraint.bound 2.3 >>> penalty = rr.nuclear_norm((5,4), lagrange=2.3) >>> penalty.bound = 3.4 Traceback (most recent call last): ... AttributeError: atom is in lagrange mode
-
set_lagrange(lagrange)¶ Set method of the lagrange property.
>>> import regreg.api as rr >>> penalty = rr.nuclear_norm((5,4), lagrange=3.4) >>> penalty.lagrange 3.4 >>> penalty.lagrange = 2.3 >>> penalty.lagrange 2.3 >>> constraint = rr.nuclear_norm((5,4), bound=3.4) >>> constraint.lagrange = 3.4 Traceback (most recent call last): ... AttributeError: atom is in bound mode
-
set_offset(value)¶
-
set_quadratic(quadratic)¶ Set the quadratic part of the composite.
-
classmethod
shift(offset, lagrange=None, diag=False, bound=None, quadratic=None)¶
-
smooth_objective(x, mode='both', check_feasibility=False)¶ The zero function.
-
smoothed(smoothing_quadratic)¶ Add quadratic smoothing term
-
solve(quadratic=None, return_optimum=False, **fit_args)¶
-
svd_tol= 1e-05¶
-
tol= 1e-05¶
-
operator_norm¶
-
class
regreg.affine.factored_matrix.operator_norm(shape, lagrange=None, bound=None, offset=None, quadratic=None, initial=None, initial_rank=10, warm_start=None)¶ Bases:
regreg.atoms.svd_norms.operator_norm-
__init__(shape, lagrange=None, bound=None, offset=None, quadratic=None, initial=None, initial_rank=10, warm_start=None)¶ Initialize self. See help(type(self)) for accurate signature.
-
classmethod
affine(linear_operator, offset, lagrange=None, diag=False, bound=None, quadratic=None)¶ This is the same as the linear class method but with offset as a positional argument
-
apply_offset(x)¶ If self.offset is not None, return x-self.offset, else return x.
-
property
bound¶
-
bound_prox(X, bound=None)¶ Return unique minimizer
\[{B}^{\delta}(\theta) = \text{argmin}_{B \in \mathbb{R}^{{n \times p}}} \frac{1}{2} \|\theta-B\|^2_2 \ \text{s.t.} \ \|B\|_{\text{op}} \leq \delta\]where \(\delta\) is the bound parameter and \(\theta\) is arg.
If the argument bound is None and the atom is in bound mode,
self.boundis used as the bound parameter, else an exception is raised.The class atom’s bound_prox just returns the appropriate bound parameter for use by the subclasses.
- Parameters
arg : np.ndarray(np.float)
Argument of the proximal map.
bound : float (optional)
Bound for the constraint on the seminorm. Defaults to self.bound.
- Returns
Z : np.ndarray(np.float)
The proximal map of arg.
-
static
check_subgradient(atom, prox_center)¶ For a given seminorm, verify the KKT condition for the problem for the proximal problem
\[\text{minimize}_u \frac{1}{2} \|u-z\|^2_2 + h(z)\]where \(z\) is the prox_center and \(h\) is atom which may be in Lagrange or bound form.
If the atom is in Lagrange form, this function should return two values equal to the seminorm of the minimizer. If it is bound form it should return two values equal to the dual seminorm of the residual, i.e. the prox_center minus the minimizer.
- Parameters
atom : seminorm
prox_center : np.ndarray(np.float)
Center for the proximal map.
- Returns
v1, v2 : float
Two values that should be equal if the proximal map is correct.
-
property
conjugate¶ The conjugate of an atom.
-
constraint(X, bound=None)¶ Verify \(\|B\|_{\text{op}} \leq \delta\), where \(\delta\) is bound.
If the result is True, returns 0, else returns np.inf.
The class seminorm’s constraint just returns the appropriate bound parameter for use by the subclasses.
-
property
dual¶
-
get_bound()¶ Get method of the bound property.
>>> import regreg.api as rr >>> constraint = rr.operator_norm((5,4), bound=2.3) >>> constraint.bound 2.3
-
get_conjugate()¶ Return the conjugate of an given atom.
>>> import regreg.api as rr >>> penalty = rr.operator_norm((5,4), lagrange=3.4) >>> penalty.get_conjugate() nuclear_norm(..., bound=3.4...)
-
get_dual()¶ Return the dual of an atom. This dual is formed by making introducing new variables \(v=Ax\) where \(A\) is self.linear_transform.
>>> import regreg.api as rr >>> penalty = rr.operator_norm((5,4), lagrange=2.3) >>> penalty operator_norm(..., lagrange=2.3...) >>> penalty.dual (<regreg.affine.identity object at 0x...>, nuclear_norm(..., bound=2.3...))
If there is a linear part to the penalty, the linear_transform may not be identity. For example, the 1D fused LASSO penalty:
>>> D = (np.identity(4) + np.diag(-np.ones(3),1))[:-1] >>> D array([[ 1., -1., 0., 0.], [ 0., 1., -1., 0.], [ 0., 0., 1., -1.]]) >>> linear_atom = rr.l1norm.linear(D, lagrange=2.3) >>> linear_atom affine_atom(l1norm((3,), lagrange=2.3...), array([[ 1., -1., 0., 0.], [ 0., 1., -1., 0.], [ 0., 0., 1., -1.]])) >>> linear_atom.dual (<regreg.affine.linear_transform object at 0x...>, supnorm((3,), bound=2.3...))
-
get_lagrange()¶ Get method of the lagrange property.
>>> import regreg.api as rr >>> penalty = rr.operator_norm((5,4), lagrange=3.4) >>> penalty.lagrange 3.4
-
get_offset()¶
-
get_quadratic()¶ Get the quadratic part of the composite.
-
property
lagrange¶
-
lagrange_prox(X, lipschitz=1, lagrange=None)¶ Return unique minimizer
\[{B}^{\lambda}(\theta) = \text{argmin}_{B \in \mathbb{R}^{{n \times p}}} \frac{L}{2} \|\theta-B\|^2_2 + \lambda \|B\|_{\text{op}} \]Above, \(\lambda\) is the Lagrange parameter and \(L\) is the Lipschitz parameter and \(\theta\) is arg.
If the argument lagrange is None and the atom is in lagrange mode,
self.lagrangeis used as the lagrange parameter, else an exception is raised.The class atom’s lagrange_prox just returns the appropriate lagrange parameter for use by the subclasses.
- Parameters
arg : np.ndarray(np.float)
Argument of the proximal map.
lipschitz : float
Coefficient in front of the quadratic.
lagrange : float (optional)
Lagrange factor in front of the seminorm. Defaults to self.lagrange.
- Returns
Z : np.ndarray(np.float)
The proximal map of arg.
-
latexify(var=None, idx='')¶ Return a LaTeX representation of an object.
>>> import regreg.api as rr >>> penalty = rr.l1norm(10, lagrange=0.9) >>> penalty.latexify(var=r'\gamma') '\\lambda_{} \\|\\gamma\\|_1'
- Parameters
var : string
Argument of the functions
idx : string
Optional subscript index.
- Returns
L : string
A LaTeX representation of the atom.
-
classmethod
linear(linear_operator, lagrange=None, diag=False, bound=None, quadratic=None, offset=None)¶
-
property
linear_transform¶ The linear transform applied before a penalty is computed. Defaults to regreg.affine.identity
>>> from regreg.api import l1norm >>> penalty = l1norm(30, lagrange=3.4) >>> type(penalty.linear_transform) <class 'regreg.affine.identity'>
-
nonsmooth_objective(arg, check_feasibility=False)¶ The nonsmooth objective function of the atom. Includes self.quadratic.objective(arg).
>>> import regreg.api as rr >>> penalty = rr.l1norm(4, lagrange=2) >>> penalty.nonsmooth_objective([3, 4, 5, 9]) 42.0 >>> 2 * sum([3, 4, 5, 9]) 42
- Parameters
arg : np.ndarray(np.float)
Argument of the seminorm.
check_feasibility : bool
If True, then return np.inf if appropriate.
- Returns
value : np.float
The seminorm of arg.
-
objective(x, check_feasibility=False)¶
-
objective_template= '\\|%(var)s\\|_{\\text{op}}'¶
-
objective_vars= {'dualnormklass': 'nuclear_norm', 'initargs': '(5,4)', 'linear': 'D', 'normklass': 'operator_norm', 'offset': '\\alpha', 'shape': '{n \\times p}', 'var': 'B'}¶
-
property
offset¶
-
prox_tol= 1e-05¶
-
proximal(quadratic, prox_control=None)¶ The proximal operator. If the atom is in Lagrange mode, this has the form
\[B^{\lambda}(\theta) = \text{argmin}_{B \in \mathbb{R}^{{n \times p}}} \frac{L}{2} \|\theta-B\|^2_2 + \lambda h(B-\alpha) + \langle B, \eta \rangle\]where \(\alpha\) is self.offset, \(\eta\) is quadratic.linear_term, \(\theta\) is quadratic.center and
\[h(B) = \|B\|_{\text{op}}\]If the atom is in bound mode, then this has the form
\[B^{\delta}(\theta) = \text{argmin}_{B \in \mathbb{R}^{{n \times p}}} \frac{L}{2} \|\theta-B\|^2_2 + \langle B, \eta \rangle \ \text{s.t.} \ h(B - \alpha) \leq \delta\]>>> import regreg.api as rr >>> penalty = rr.l1norm(4, lagrange=2) >>> Q = rr.identity_quadratic(1.5, [3, -4, -1, 1], 0, 0) >>> penalty.proximal(Q) array([ 1.6666..., -2.6666..., -0. , 0. ])
- Parameters
quadratic : regreg.identity_quadratic.identity_quadratic
A quadratic added to the atom before minimizing.
prox_control : [None, dict]
This argument is ignored for seminorms, but otherwise is passed to regreg.algorithms.FISTA if the atom needs to be solved iteratively.
- Returns
Z : np.ndarray(np.float)
The proximal map of the implied center of quadratic.
-
proximal_optimum(quadratic)¶
-
proximal_step(quadratic, prox_control=None)¶ Compute the proximal optimization
- Parameters
prox_control: [None, dict]
If not None, then a dictionary of parameters for the prox procedure
-
property
quadratic¶ Quadratic part of the object, instance of regreg.identity_quadratic.identity_quadratic.
-
seminorm(X, check_feasibility=False, lagrange=None)¶ Return \(\lambda \cdot \|B\|_{\text{op}}\), where \(\lambda\) is lagrange. If check_feasibility is True, and seminorm is unbounded, will return np.inf if appropriate.
The class seminorm’s seminorm just returns the appropriate lagrange parameter for use by the subclasses.
-
set_bound(bound)¶ Set method of the bound property.
>>> import regreg.api as rr >>> constraint = rr.operator_norm((5,4), bound=3.4) >>> constraint.bound 3.4 >>> constraint.bound = 2.3 >>> constraint.bound 2.3 >>> penalty = rr.operator_norm((5,4), lagrange=2.3) >>> penalty.bound = 3.4 Traceback (most recent call last): ... AttributeError: atom is in lagrange mode
-
set_lagrange(lagrange)¶ Set method of the lagrange property.
>>> import regreg.api as rr >>> penalty = rr.operator_norm((5,4), lagrange=3.4) >>> penalty.lagrange 3.4 >>> penalty.lagrange = 2.3 >>> penalty.lagrange 2.3 >>> constraint = rr.operator_norm((5,4), bound=3.4) >>> constraint.lagrange = 3.4 Traceback (most recent call last): ... AttributeError: atom is in bound mode
-
set_offset(value)¶
-
set_quadratic(quadratic)¶ Set the quadratic part of the composite.
-
classmethod
shift(offset, lagrange=None, diag=False, bound=None, quadratic=None)¶
-
smooth_objective(x, mode='both', check_feasibility=False)¶ The zero function.
-
smoothed(smoothing_quadratic)¶ Add quadratic smoothing term
-
solve(quadratic=None, return_optimum=False, **fit_args)¶
-
svd_tol= 0.0001¶
-
tol= 1e-05¶
-
Functions¶
-
regreg.affine.factored_matrix.compute_iterative_svd(transform, initial_rank=None, warm_start=None, min_singular=1e-16, tol=1e-05, debug=False, stopping_rule=None, padding=5, update_rank=<function <lambda>>, max_rank=None)¶ Compute the SVD of a matrix using partial_svd.
If no initial rank is given, it assumes a rank of size min(n,p) / 10.
Iteratively calls
partial_svd()until the singular_values are small enough.- Parameters
transform : [linear_transform, ndarray]
Linear_transform whose SVD is computed. If an ndarray, it is first cast with
astransform()initial_rank : None or int, optional
A guess at the rank of the matrix.
warm_start : np.ndarray(np.float), optional
A guess at the left singular vectors of the matrix. For fat matrices, these should be right singular vectors, while for tall matrices these should be left singular vectors.
min_singular : np.float, optional
Stop when the singular value has this relative tolerance.
tol: np.float, optional
Tolerance at which the norm of the singular values are deemed to have converged.
debug: bool, optional
Print debugging statements.
stopping_rule : None or callable, optional
Function of the singular values
D, returning True | False, used to determine whether to stop. Continue whilestopping_rule(D) == False, or when stopping_rule is None.padding : int, optional
How many more singular vectors are found.
update_rank : callable, optional
A rule to update rank, defaults to doubling the rank.
max_rank : None or int, optional
Largest rank considered. Defaults to 2 * min(transform.output_shape[0], transform.input_shape[0]). An exception is raised if algorithm exceeds given value.
- Returns
U, D, VT, Ufull : np.ndarray(np.float)
An SVD of the transform. Ufull is the full set of left singular vectors found.
Examples
>>> import regreg.api as rr >>> np.random.seed(0) >>> X = np.random.standard_normal((100, 200)) >>> U, D, VT = rr.compute_iterative_svd(X)[:3] >>> np.linalg.norm(np.dot(U.T, U) - np.identity(100)) < 1.e-6 True
>>> np.linalg.norm(np.dot(VT, VT.T) - np.identity(100)) < 1.e-6 True
-
regreg.affine.factored_matrix.partial_svd(transform, rank=1, padding=5, max_its=1000, tol=1e-08, warm_start=None, return_full=False, debug=False, stopping_rule=None)¶ Compute partial SVD of the linear transform transform
Uses the Mazumder/Hastie algorithm in (TODO: CITE)
- Parameters
transform : [linear_transform, ndarray]
Linear_transform whose SVD is computed. If an ndarray, it is first cast with
astransform()rank : int, optional
What rank of SVD to compute.
padding : int, optional
Compute a few further singular values / vectors. This results in a better estimator of the rank of interest.
max_its : int, optional
How many iterations of the full cycle to complete.
tol : np.float, optional
Tolerance at which the norm of the singular values are deemed to have converged.
warm_start : np.ndarray(np.float), optional
A guess at the singular vectors of the matrix. For fat matrices, these should be right singular vectors, while for tall matrices these should be left singular vectors.
return_full: bool, optional
Return a singular values / vectors from padding?
debug: bool, optional
Print debugging statements.
stopping_rule : callable
Function of the singular values used to determine whether to stop.
- Returns
U, D, VT, Ufull : np.ndarray(np.float)
An SVD up to rank of the transform. Ufull is the full set of left singular vectors found.
Examples
>>> from regreg.affine.factored_matrix import partial_svd >>> np.random.seed(0) >>> X = np.random.standard_normal((100, 200)) >>> U, D, VT = partial_svd(X, rank=10)[:3] >>> assert(np.linalg.norm(np.dot(U.T, U) - np.identity(10)) < 1.e-4) >>> assert(np.linalg.norm(np.dot(VT, VT.T) - np.identity(10)) < 1.e-4) >>> U_np, D_np, VT_np = np.linalg.svd(X, full_matrices=False) >>> assert(np.linalg.norm(U - np.dot(U, np.dot(U_np.T[:10], U_np[:,:10]))) < 1.e-4) >>> assert(np.linalg.norm(VT - np.dot(VT, np.dot(VT_np[:10].T, VT_np[:10]))) < 1.e-4)