"""
Copyright 2013 Steven Diamond
This file is part of CVXPY.
CVXPY is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.
CVXPY is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with CVXPY. If not, see <http://www.gnu.org/licenses/>.
"""
from cvxpy.atoms.atom import Atom
from cvxpy.atoms.affine.index import index
import numpy as np
import scipy.sparse as sp
from ..utilities.power_tools import (fracify, decompose, approx_error, lower_bound,
over_bound, prettydict, gm_constrs)
import cvxpy.lin_ops.lin_utils as lu
[docs]class geo_mean(Atom):
""" The (weighted) geometric mean of vector ``x``, with optional powers given by ``p``:
.. math::
\\left(x_1^{p_1} \cdots x_n^{p_n} \\right)^{\\frac{1}{\mathbf{1}^Tp}}
The powers ``p`` can be a ``list``, ``tuple``, or ``numpy.array`` of nonnegative
``int``, ``float``, or ``Fraction`` objects with nonzero sum.
If not specified, ``p`` defaults to a vector of all ones, giving the
**unweighted** geometric mean
.. math::
x_1^{1/n} \cdots x_n^{1/n}.
The geometric mean includes an implicit constraint that :math:`x_i \geq 0`
whenever :math:`p_i > 0`. If :math:`p_i = 0`, :math:`x_i` will be unconstrained.
The only exception to this rule occurs when
``p`` has exactly one nonzero element, say, ``p_i``, in which case
``geo_mean(x, p)`` is equivalent to ``x_i`` (without the nonnegativity constraint).
A specific case of this is when :math:`x \in \mathbf{R}^1`.
.. note::
Generally, ``p`` cannot be represented exactly, so a rational,
i.e., fractional, **approximation** must be made.
Internally, ``geo_mean`` immediately computes an approximate normalized
weight vector :math:`w \\approx p/\mathbf{1}^Tp`
and the ``geo_mean`` atom is represented as
.. math::
x_1^{w_1} \cdots x_n^{w_n},
where the elements of ``w`` are ``Fraction`` objects that sum to exactly 1.
The maximum denominator used in forming the rational approximation
is given by ``max_denom``, which defaults to 1024, but can be adjusted
to modify the accuracy of the approximations.
The approximating ``w`` and the approximation error can be
found through the attributes ``geo_mean.w`` and ``geo_mean.approx_error``.
Examples
--------
The weights ``w`` can be seen from the string representation of the
``geo_mean`` object, or through the ``w`` attribute.
>>> from cvxpy import Variable, geo_mean, Problem, Maximize
>>> x = Variable(3, name='x')
>>> print(geo_mean(x))
geo_mean(x, (1/3, 1/3, 1/3))
>>> g = geo_mean(x, [1, 2, 1])
>>> g.w
(Fraction(1, 4), Fraction(1, 2), Fraction(1, 4))
Floating point numbers with few decimal places can sometimes be represented
exactly. The approximation error between ``w`` and ``p/sum(p)`` is given by
the ``approx_error`` attribute.
>>> import numpy as np
>>> x = Variable(4, name='x')
>>> p = np.array([.12, .34, .56, .78])
>>> g = geo_mean(x, p)
>>> g.w
(Fraction(1, 15), Fraction(17, 90), Fraction(14, 45), Fraction(13, 30))
>>> g.approx_error
0.0
In general, the approximation is not exact.
>>> p = [.123, .456, .789, .001]
>>> g = geo_mean(x, p)
>>> g.w
(Fraction(23, 256), Fraction(341, 1024), Fraction(295, 512), Fraction(1, 1024))
>>> 1e-4 <= g.approx_error <= 1e-3
True
The weight vector ``p`` can contain combinations of ``int``, ``float``,
and ``Fraction`` objects.
>>> from fractions import Fraction
>>> x = Variable(4, name='x')
>>> g = geo_mean(x, [.1, Fraction(1,3), 0, 2])
>>> print(g)
geo_mean(x, (3/73, 10/73, 0, 60/73))
>>> g.approx_error <= 1e-10
True
Sequences of ``Fraction`` and ``int`` powers can often be represented **exactly**.
>>> p = [Fraction(1,17), Fraction(4,9), Fraction(1,3), Fraction(25,153)]
>>> x = Variable(4, name='x')
>>> print(geo_mean(x, p))
geo_mean(x, (1/17, 4/9, 1/3, 25/153))
Terms with a zero power will not have an implicit nonnegativity constraint.
>>> p = [1, 0, 1]
>>> x = Variable(3, name='x')
>>> obj = Maximize(geo_mean(x,p))
>>> constr = [sum(x) <= 1, -1 <= x, x <= 1]
>>> val = Problem(obj, constr).solve()
>>> x = np.array(x.value).flatten()
>>> print(x)
[ 1. -1. 1.]
Parameters
----------
x : cvxpy.Variable
A column or row vector whose elements we will take the geometric mean of.
p : Sequence (list, tuple, ...) of ``int``, ``float``, or ``Fraction`` objects
A vector of weights for the weighted geometric mean
When ``p`` is a sequence of ``int`` and/or ``Fraction`` objects,
``w`` can often be an **exact** representation of the weights.
An exact representation is sometimes possible when ``p`` has ``float``
elements with only a few decimal places.
max_denom : int
The maximum denominator to use in approximating ``p/sum(p)`` with
``geo_mean.w``. If ``w`` is not an exact representation, increasing
``max_denom`` **may** offer a more accurate representation, at the
cost of requiring more convex inequalities to represent the geometric mean.
Attributes
----------
w : tuple of ``Fractions``
A rational approximation of ``p/sum(p)``.
approx_error : float
The error in approximating ``p/sum(p)`` with ``w``, given by
:math:`\|p/\mathbf{1}^T p - w \|_\infty`
"""
def __init__(self, x, p=None, max_denom=1024):
""" Implementation details of geo_mean.
Attributes
----------
w_dyad : tuple of ``Fractions`` whose denominators are all a power of two
The dyadic completion of ``w``, which is used internally to form the
inequalities representing the geometric mean.
tree : ``dict``
keyed by dyadic tuples, whose values are Sequences of children.
The children are also dyadic tuples.
This represents the graph that needs to be formed to represent the
weighted geometric mean.
cone_lb : int
A known lower bound (which is not always tight) on the number of cones
needed to represent this geometric mean.
cone_num_over : int
The number of cones beyond the lower bound that this geometric mean used.
If 0, we know that it used the minimum possible number of cones.
Since cone_lb is not always tight, it may be using the minimum number of cones even if
cone_num_over is not 0.
cone_num : int
The number of second order cones used to form this geometric mean
"""
super(geo_mean, self).__init__(x)
x = self.args[0]
if x.shape[0] == 1:
n = x.shape[1]
elif x.shape[1] == 1:
n = x.shape[0]
else:
raise ValueError('x must be a row or column vector.')
if p is None:
p = [1]*n
if len(p) != n:
raise ValueError('x and p must have the same number of elements.')
if any(v < 0 for v in p) or sum(p) <= 0:
raise ValueError('powers must be nonnegative and not all zero.')
self.w, self.w_dyad = fracify(p, max_denom)
self.approx_error = approx_error(p, self.w)
self.tree = decompose(self.w_dyad)
# known lower bound on number of cones needed to represent w_dyad
self.cone_lb = lower_bound(self.w_dyad)
# number of cones used past known lower bound
self.cone_num_over = over_bound(self.w_dyad, self.tree)
# number of cones used
self.cone_num = self.cone_lb + self.cone_num_over
# Returns the (weighted) geometric mean of the elements of x.
@Atom.numpy_numeric
def numeric(self, values):
values = np.array(values[0]).flatten()
val = 1.0
for x, p in zip(values, self.w):
val *= x**float(p)
return val
def _domain(self):
"""Returns constraints describing the domain of the node.
"""
# No special case when only one non-zero weight.
selection = np.array([w_i > 0 for w_i in self.w])
return [self.args[0][selection > 0] >= 0]
def _grad(self, values):
"""Gives the (sub/super)gradient of the atom w.r.t. each argument.
Matrix expressions are vectorized, so the gradient is a matrix.
Args:
values: A list of numeric values for the arguments.
Returns:
A list of SciPy CSC sparse matrices or None.
"""
x = np.matrix(values[0])
# No special case when only one non-zero weight.
w_arr = np.array([float(w_i) for w_i in self.w])
# Outside domain.
if np.any(x[w_arr > 0] <= 0):
return [None]
else:
D = w_arr/x.A.ravel(order='F')*self.numeric(values)
return [sp.csc_matrix(D).T]
def name(self):
return "%s(%s, (%s))" % (self.__class__.__name__,
self.args[0].name(),
', '.join(str(v) for v in self.w))
def pretty_tree(self):
print(prettydict(self.tree))
def shape_from_args(self):
"""Returns the (row, col) shape of the expression.
"""
return (1, 1)
def sign_from_args(self):
"""Returns sign (is positive, is negative) of the expression.
"""
# Always positive.
return (True, False)
def is_atom_convex(self):
"""Is the atom convex?
"""
return False
def is_atom_concave(self):
"""Is the atom concave?
"""
return True
def is_incr(self, idx):
"""Is the composition non-decreasing in argument idx?
"""
return True
def is_decr(self, idx):
"""Is the composition non-increasing in argument idx?
"""
return False
def validate_arguments(self):
# since correctly validating arguments with this function is tricky,
# we do it in __init__ instead.
pass
def get_data(self):
return [self.w, self.w_dyad, self.tree]
def copy(self, args=None):
"""Returns a shallow copy of the geo_mean atom.
Parameters
----------
args : list, optional
The arguments to reconstruct the atom. If args=None, use the
current args of the atom.
Returns
-------
geo_mean atom
"""
if args is None:
args = self.args
# Avoid calling __init__() directly as we do not have p and max_denom.
copy = type(self).__new__(type(self))
super(type(self), copy).__init__(*args)
# Emulate __init__()
copy.w, copy.w_dyad, copy.tree = self.get_data()
copy.approx_error = self.approx_error
copy.cone_lb = self.cone_lb
copy.cone_num_over = self.cone_num_over
copy.cone_num = self.cone_num
return copy
@staticmethod
def graph_implementation(arg_objs, shape, data=None):
"""Reduces the atom to an affine expression and list of constraints.
Parameters
----------
arg_objs : list
LinExpr for each argument.
shape : tuple
The shape of the resulting expression.
data :
Additional data required by the atom.
Returns
-------
tuple
(LinOp for objective, list of constraints)
"""
w, w_dyad, tree = data
t = lu.create_var((1, 1))
if arg_objs[0].shape[1] == 1:
x_list = [index.get_index(arg_objs[0], [], i, 0) for i in range(len(w))]
if arg_objs[0].shape[0] == 1:
x_list = [index.get_index(arg_objs[0], [], 0, i) for i in range(len(w))]
# todo: catch cases where we have (0, 0, 1)?
# todo: what about curvature case (should be affine) in trivial case of (0, 0 , 1),
# should this behavior match with what we do in power?
return t, gm_constrs(t, x_list, w)
