# coding: utf-8
# ------------------------------------------------------------------------------
# Copyright (C) 2019 Maximilian Stahlberg
#
# This file is part of PICOS.
#
# PICOS 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.
#
# PICOS 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 this program. If not, see <http://www.gnu.org/licenses/>.
# ------------------------------------------------------------------------------
"""Quadratic constraint types."""
from collections import namedtuple
from .. import glyphs
from ..apidoc import api_end, api_start
from ..caching import cached_property
from .constraint import Constraint, ConstraintConversion
_API_START = api_start(globals())
# -------------------------------
[docs]class NonconvexQuadraticConstraint(Constraint):
"""Bound on a nonconvex quadratic expression."""
[docs] def __init__(self, lhs, relation, rhs):
"""Construct a :class:`NonconvexQuadraticConstraint`.
:param lhs: Left hand side quadratic or affine expression.
:type lhs: ~picos.expressions.QuadraticExpression or
~picos.expressions.AffineExpression
:param str relation: Constraint relation symbol.
:param rhs: Right hand side quadratic or affine expression.
:type rhs: ~picos.expressions.QuadraticExpression or
~picos.expressions.AffineExpression
"""
from ..expressions import AffineExpression, QuadraticExpression
assert isinstance(lhs, (AffineExpression, QuadraticExpression))
assert isinstance(rhs, (AffineExpression, QuadraticExpression))
assert any(isinstance(exp, QuadraticExpression) for exp in (lhs, rhs))
assert relation in self.LE + self.GE
assert lhs.size == rhs.size
self.lhs = lhs
self.rhs = rhs
self.relation = relation
Constraint.__init__(self, self._get_type_term())
def _get_type_term(self):
return "Nonconvex Quadratic"
# TODO: Add common interface for all the le0, ge0, smaller, greater methods?
@cached_property
def le0(self):
"""Quadratic expression constrained to be at most zero."""
if self.relation == self.GE:
return self.rhs - self.lhs
else:
return self.lhs - self.rhs
@property
def smaller(self):
"""Smaller-or-equal side of the constraint."""
return self.rhs if self.relation == self.GE else self.lhs
@property
def greater(self):
"""Greater-or-equal side of the constraint."""
return self.lhs if self.relation == self.GE else self.rhs
Subtype = namedtuple("Subtype", ("qterms"))
def _subtype(self):
return self.Subtype(self.le0.num_quad_terms)
@classmethod
def _cost(cls, subtype):
return subtype.qterms
def _expression_names(self):
yield "lhs"
yield "rhs"
def _str(self):
if self.relation == self.LE:
return glyphs.le(self.lhs.string, self.rhs.string)
else:
return glyphs.ge(self.lhs.string, self.rhs.string)
def _get_size(self):
return (1, 1)
def _get_slack(self):
if self.relation == self.LE:
return self.rhs.value - self.lhs.value
else:
return self.lhs.value - self.rhs.value
[docs]class ConicQuadraticConstraint(NonconvexQuadraticConstraint):
r"""Bound on a *nearly* convex quadratic expression.
Nearly convex means that the bilinear form representing the quadratic part
of the expression that the constraint poses to be at most zero has exactly
one negative eigenvalue. More precisely, if the constraint is of the form
:math:`x^TQx + a^Tx + b \leq x^TRx + b^Tx + c`, then `Q - R` needs to have
exactly one negative eigenvalue for the constraint to be nearly convex. Such
constraints can be posed as conic constraints under an additional assumption
that some affine term is nonnegative.
At this point, this class may only be used for nonconvex constraints of the
form :math:`x^TQx + p^Tx + q \leq (a^Tx + b)(c^Tx + d)` with
:math:`x^TQx + p^Tx + q` representable as a squared norm. In this case, the
additional assumptions required for a conic reformulation are
:math:`a^Tx + b \geq 0` and :math:`b^Tx + c \geq 0`.
Whether a constraint of this type is strengthened to a conic constraint or
relabeled as a :class:`NonconvexQuadraticConstraint` depends on the
:ref:`assume_conic <option_assume_conic>` option.
:Example:
>>> from picos import Options, RealVariable
>>> x, y, z = RealVariable("x"), RealVariable("y"), RealVariable("z")
>>> C = x**2 + 1 <= y*z; C
<Conic Quadratic Constraint: x² + 1 ≤ y·z>
>>> P = C.__class__.Conversion.convert(C, Options(assume_conic=True))
>>> list(P.constraints.values())[0]
<4×1 RSOC Constraint: ‖fullroot(x² + 1)‖² ≤ y·z ∧ y ≥ 0>
>>> Q = C.__class__.Conversion.convert(C, Options(assume_conic=False))
>>> list(Q.constraints.values())[0]
<Nonconvex Quadratic Constraint: x² + 1 ≤ y·z>
.. note::
Solver implementations must not support this constraint type so that
the user's choice for the :ref:`assume_conic <option_assume_conic>`
option is respected.
"""
[docs] class Conversion(ConstraintConversion):
"""Nearly convex quadratic to (rotated) second order cone conversion."""
[docs] @classmethod
def predict(cls, subtype, options):
"""Implement :meth:`~.constraint.ConstraintConversion.predict`."""
from . import AbsoluteValueConstraint, SOCConstraint, RSOCConstraint
if options.assume_conic:
n = subtype.conic_argdim
if subtype.rotated:
yield ("con", RSOCConstraint.make_type(argdim=n), 1)
elif n == 1:
yield ("con", AbsoluteValueConstraint.make_type(), 1)
else:
yield ("con", SOCConstraint.make_type(argdim=n), 1)
else:
yield ("con", NonconvexQuadraticConstraint.make_type(), 1)
[docs] @classmethod
def convert(cls, con, options):
"""Implement :meth:`~.constraint.ConstraintConversion.convert`."""
from ..expressions import AffineExpression
from ..modeling import Problem
from . import AbsoluteValueConstraint, SOCConstraint, RSOCConstraint
P = Problem()
if options.assume_conic:
q, r = con.smaller, con.greater
root = AffineExpression.zero() if q.is0 else q.fullroot
if r.scalar_factors[0] is not r.scalar_factors[1]:
P.add_constraint(RSOCConstraint(root, *r.scalar_factors))
else:
bound = r.scalar_factors[0]
if len(root) == 1:
P.add_constraint(AbsoluteValueConstraint(root, bound))
else:
P.add_constraint(SOCConstraint(root, bound))
else:
P.add_constraint(NonconvexQuadraticConstraint(
con.lhs, con.relation, con.rhs))
return P
[docs] def __init__(self, lhs, relation, rhs):
"""Construct a :class:`ConicQuadraticConstraint`.
See :meth:`NonconvexQuadraticConstraint.__init__` for more.
"""
from ..expressions import QuadraticExpression
NonconvexQuadraticConstraint.__init__(self, lhs, relation, rhs)
# Additional usage restrictions for the near-convex case.
assert isinstance(self.lhs, QuadraticExpression)
assert isinstance(self.rhs, QuadraticExpression)
assert self.smaller.is_squared_norm
assert self.greater.scalar_factors
def _get_type_term(self):
return "Conic Quadratic"
Subtype = namedtuple("Subtype", ("qterms", "conic_argdim", "rotated"))
def _subtype(self):
qterms = self.le0.num_quad_terms
conic_argdim = 1 if self.smaller.is0 else len(self.smaller.fullroot)
sf = self.greater.scalar_factors
rotated = sf[0] is not sf[1]
return self.Subtype(qterms, conic_argdim, rotated)
[docs]class ConvexQuadraticConstraint(NonconvexQuadraticConstraint):
"""Bound on a convex quadratic expression."""
[docs] class ConicConversion(ConstraintConversion):
"""Convex quadratic to (rotated) second order cone conversion."""
[docs] @classmethod
def predict(cls, subtype, options):
"""Implement :meth:`~.constraint.ConstraintConversion.predict`."""
from . import AbsoluteValueConstraint, SOCConstraint, RSOCConstraint
if subtype.haslin:
yield ("con", RSOCConstraint.make_type(argdim=subtype.rank), 1)
elif subtype.rank == 1:
yield ("con", AbsoluteValueConstraint.make_type(), 1)
else:
yield ("con", SOCConstraint.make_type(argdim=subtype.rank), 1)
[docs] @classmethod
def convert(cls, con, options):
"""Implement :meth:`~.constraint.ConstraintConversion.convert`."""
from ..expressions import AffineExpression
from ..modeling import Problem
from . import AbsoluteValueConstraint, SOCConstraint, RSOCConstraint
le0 = con.le0
P = Problem()
if con.subtype.haslin:
one = AffineExpression.from_constant(1)
P.add_constraint(RSOCConstraint(le0.quadroot, -le0._aff, one))
else:
assert le0._aff.constant
value = -le0._aff.value
if value < 0:
raise ValueError("The constraint {} is infeasible as it "
"upper-bounds a positive semidefinite quadratic form by"
" a negative constant.".format(con))
root = le0.quadroot
bound = AffineExpression.from_constant(value**0.5,
name=glyphs.sqrt(glyphs.scalar(value)))
if len(root) == 1:
P.add_constraint(AbsoluteValueConstraint(root, bound))
else:
P.add_constraint(SOCConstraint(root, bound))
return P
[docs] def __init__(self, lhs, relation, rhs):
"""Construct a :class:`ConvexQuadraticConstraint`.
See :meth:`NonconvexQuadraticConstraint.__init__` for more.
"""
NonconvexQuadraticConstraint.__init__(self, lhs, relation, rhs)
def _get_type_term(self):
return "Convex Quadratic"
Subtype = namedtuple("Subtype", ("qterms", "rank", "haslin"))
def _subtype(self):
from ..expressions import QuadraticExpression
# HACK: Be consistent with the prediction, which cannot foresee whether
# the linear terms of both sides cancel out.
lhs_aff_is_cst = self.lhs._aff.constant \
if isinstance(self.lhs, QuadraticExpression) else self.lhs.constant
rhs_aff_is_cst = self.rhs._aff.constant \
if isinstance(self.rhs, QuadraticExpression) else self.rhs.constant
haslin = not lhs_aff_is_cst or not rhs_aff_is_cst
return self.Subtype(self.le0.num_quad_terms, self.le0.rank, haslin)
# --------------------------------------
__all__ = api_end(_API_START, globals())