exprQuad Class Reference

class exprQuad, with constant, linear and quadratic terms More...

#include <exprQuad.hpp>

Inheritance diagram for exprQuad:

Inheritance graph
[legend]
Collaboration diagram for exprQuad:

Collaboration graph
[legend]
List of all members.

Convexification data structures

These are filled by alphaConvexify, which implements the alpha-convexification method described in the LaGO paper by Nowak and Vigerske -- see also Adjiman and Floudas.

 exprQuad (CouNumber c0, std::vector< std::pair< exprVar *, CouNumber > > &lcoeff, std::vector< quadElem > &qcoeff, expression **al=NULL, int n=0)
 Constructor.
 exprQuad (const exprQuad &src, Domain *d=NULL)
 Copy constructor.
sparseQgetQ () const
 Constructor.
int getnQTerms ()
 Constructor.
virtual expressionclone (Domain *d=NULL) const
 cloning method
virtual void print (std::ostream &=std::cout, bool=false) const
 Print expression to an iostream.
virtual CouNumber operator() ()
 Function for the evaluation of the expression.
CouNumber gradientNorm (const double *x)
 return l-2 norm of gradient at given point
virtual expressiondifferentiate (int index)
 Compute derivative of this expression with respect to variable whose index is passed as argument.
virtual expressionsimplify ()
 Simplify expression.
virtual int Linearity ()
 Get a measure of "how linear" the expression is.
virtual void getBounds (expression *&, expression *&)
 Get lower and upper bound of an expression (if any).
virtual void getBounds (CouNumber &, CouNumber &)
 Get lower and upper bound of an expression (if any).
virtual void generateCuts (expression *w, OsiCuts &cs, const CouenneCutGenerator *cg, t_chg_bounds *=NULL, int=-1, CouNumber=-COUENNE_INFINITY, CouNumber=COUENNE_INFINITY)
 Generate cuts for the quadratic expression, which are supporting hyperplanes of the concave upper envelope and the convex lower envelope.
virtual bool alphaConvexify (const CouenneProblem *)
 Compute data for $\alpha$-convexification of a quadratic form (fills in dCoeff_ and dIndex_ for the convex underestimator).
void quadCuts (expression *w, OsiCuts &cs, const CouenneCutGenerator *cg)
 method exprQuad::quadCuts Based on the information (dIndex_, dCoeffLo_, dCoeffUp_) created/modified by alphaConvexify(), create convexification cuts for this expression.
virtual int compare (exprQuad &)
 Compare two exprQuad.
virtual enum expr_type code ()
 Code for comparisons.
virtual int rank ()
 Used in rank-based branching variable choice.
virtual bool isInteger ()
 is this expression integer?
virtual int DepList (std::set< int > &deplist, enum dig_type type=ORIG_ONLY)
 fill in the set with all indices of variables appearing in the expression
virtual CouNumber selectBranch (const CouenneObject *obj, const OsiBranchingInformation *info, expression *&var, double *&brpts, double *&brDist, int &way)
 Set up branching object by evaluating many branching points for each expression's arguments.
virtual void fillDepSet (std::set< DepNode *, compNode > *dep, DepGraph *g)
 Fill dependence set of the expression associated with this auxiliary variable.
virtual void replace (exprVar *x, exprVar *w)
 replace variable x with new (aux) w
virtual void realign (const CouenneProblem *p)
 replace variable x with new (aux) w
virtual bool impliedBound (int, CouNumber *, CouNumber *, t_chg_bounds *)
 implied bound processing
CouNumber computeQBound (int sign)
 method to compute the bound based on sign: -1 for lower, +1 for upper
virtual void closestFeasible (expression *varind, expression *vardep, CouNumber &left, CouNumber &right) const
 compute $y^{lv}$ and $y^{uv}$ for Violation Transfer algorithm
void computeQuadFiniteBound (CouNumber &qMin, CouNumber &qMax, CouNumber *l, CouNumber *u, int &indInfLo, int &indInfUp)
 return lower and upper bound of quadratic expression
virtual bool isCuttable (CouenneProblem *problem, int index) const
 can this expression be further linearized or are we on its concave ("bad") side
std::vector< std::pair< CouNumber,
std::vector< std::pair< exprVar *,
CouNumber > > > > 
eigen_
 eigenvalues and eigenvectors
std::map< exprVar *, std::pair<
CouNumber, CouNumber > > 
bounds_
 current bounds (checked before re-computing eigenvalues/vectors)
int nqterms_
 number of non-zeroes in Q

Public Types

typedef std::vector< std::pair<
exprVar *, CouNumber > > 
sparseQcol
 matrix
typedef std::vector< std::pair<
exprVar *, sparseQcol > > 
sparseQ

Protected Attributes

Q matrix storage
Sparse implementation: given expression of the form $\sum_{i \in N, j \in N} q_{ij} x_i x_j$, qindexI_ and qindexJ_ contain respectively entries $i$ and $j$ for which $q_{ij}$ is nonzero in $q_{ij} x_i x_j$:

sparseQ matrix_

Detailed Description

class exprQuad, with constant, linear and quadratic terms

It represents an expression of the form $a_0 + \sum_{i\in I} b_i x_i + x^T Q x + \sum_{i \in J} h_i (x)$, with $a_0 + \sum_{i\in I} b_i x_i$ an affine term, $x^T Q x$ a quadratic term, and a nonlinear sum $\sum_{i \in J} h_i (x)$. Standardization checks possible quadratic or linear terms in the latter and includes them in the former parts.

If $h_i(x)$ is a product of two nonlinear, nonquadratic functions $h'(x)h''(x)$, two auxiliary variables $w'=f'(x)$ and $w''=h''(x)$ are created and the product $w'w''$ is included in the quadratic part of the exprQuad. If $h(x)$ nonquadratic, nonlinear function, an auxiliary variable $w=h(x)$ is created and included in the linear part.

Definition at line 42 of file exprQuad.hpp.


Member Typedef Documentation

typedef std::vector<std::pair <exprVar *, CouNumber> > exprQuad::sparseQcol

matrix

Definition at line 47 of file exprQuad.hpp.

typedef std::vector<std::pair <exprVar *, sparseQcol> > exprQuad::sparseQ

Definition at line 48 of file exprQuad.hpp.


Constructor & Destructor Documentation

exprQuad::exprQuad ( CouNumber  c0,
std::vector< std::pair< exprVar *, CouNumber > > &  lcoeff,
std::vector< quadElem > &  qcoeff,
expression **  al = NULL,
int  n = 0 
)

Constructor.

Definition at line 30 of file exprQuad.cpp.

References bounds_, COUENNE_EPS, expression::Index(), matrix_, and nqterms_.

Referenced by clone().

exprQuad::exprQuad ( const exprQuad src,
Domain d = NULL 
)

Copy constructor.

Definition at line 129 of file exprQuad.cpp.

References clone(), eigen_, expression::Index(), and matrix_.


Member Function Documentation

sparseQ& exprQuad::getQ (  )  const [inline]

Constructor.

Definition at line 92 of file exprQuad.hpp.

References matrix_.

int exprQuad::getnQTerms (  )  [inline]

Constructor.

Definition at line 95 of file exprQuad.hpp.

References nqterms_.

virtual expression* exprQuad::clone ( Domain d = NULL  )  const [inline, virtual]

cloning method

Reimplemented from exprGroup.

Definition at line 99 of file exprQuad.hpp.

References exprQuad().

Referenced by exprQuad().

void exprQuad::print ( std::ostream &  = std::cout,
bool  = false 
) const [virtual]

Print expression to an iostream.

Reimplemented from exprGroup.

Definition at line 172 of file exprQuad.cpp.

References COUENNE_EPS, expression::Index(), m, matrix_, MAX_ARG_LINE, n, and exprGroup::print().

Referenced by impliedBound(), and quadCuts().

CouNumber exprQuad::operator() (  )  [inline, virtual]

Function for the evaluation of the expression.

Reimplemented from exprGroup.

Definition at line 291 of file exprQuad.hpp.

References expression::Index(), matrix_, exprGroup::operator()(), and x.

CouNumber exprQuad::gradientNorm ( const double *  x  )  [virtual]

return l-2 norm of gradient at given point

Reimplemented from exprGroup.

Definition at line 524 of file exprQuad.cpp.

References expression::Index(), and matrix_.

expression * exprQuad::differentiate ( int  index  )  [virtual]

Compute derivative of this expression with respect to variable whose index is passed as argument.

Reimplemented from exprGroup.

Definition at line 221 of file exprQuad.cpp.

References exprOp::arglist_, COUENNE_EPS, expression::dependsOn(), exprGroup::exprGroup(), exprSum::exprSum(), expression::Index(), exprGroup::lcoeff_, matrix_, and exprOp::nargs_.

expression * exprQuad::simplify (  )  [virtual]

Simplify expression.

Reimplemented from exprGroup.

Definition at line 541 of file exprQuad.cpp.

References exprOp::simplify().

virtual int exprQuad::Linearity (  )  [inline, virtual]

Get a measure of "how linear" the expression is.

Reimplemented from exprGroup.

Definition at line 119 of file exprQuad.hpp.

References exprGroup::c0_, CONSTANT, COUENNE_EPS, exprGroup::lcoeff_, LINEAR, exprSum::Linearity(), matrix_, QUADRATIC, and ZERO.

void exprQuad::getBounds ( expression *&  ,
expression *&   
) [virtual]

Get lower and upper bound of an expression (if any).

Reimplemented from exprGroup.

Definition at line 22 of file conv-exprQuad.cpp.

void exprQuad::getBounds ( CouNumber ,
CouNumber  
) [virtual]

Get lower and upper bound of an expression (if any).

Reimplemented from exprGroup.

Definition at line 34 of file conv-exprQuad.cpp.

References expression::getBounds().

void exprQuad::generateCuts ( expression w,
OsiCuts &  cs,
const CouenneCutGenerator cg,
t_chg_bounds = NULL,
int  = -1,
CouNumber  = -COUENNE_INFINITY,
CouNumber  = COUENNE_INFINITY 
) [virtual]

Generate cuts for the quadratic expression, which are supporting hyperplanes of the concave upper envelope and the convex lower envelope.

Reimplemented from exprGroup.

Definition at line 41 of file conv-exprQuad.cpp.

References alphaConvexify(), COUENNE_EPS, quadCuts(), and w.

bool exprQuad::alphaConvexify ( const CouenneProblem p  )  [virtual]

Compute data for $\alpha$-convexification of a quadratic form (fills in dCoeff_ and dIndex_ for the convex underestimator).

For the underestimator, dCoeffLo_ is computed such that

x^TQx + sum_i dCoeffLo_i (x_i - lb x_i)(ub x_i - x-i) is convex and underestimating (alpha_i is negative),

Regarding the overestimator, dCoeffUp_ are computed such that

x^TQx + sum_i dCoeffUp_i (x_i - lb x_i)(ub x_i - x-i) is concave and overestimating (alpha_i is positive).

If the method hasn't been called, dIndex_ will be NULL. If Q is positive semidefinite, then dCoeffLo_ will be NULL. If Q is negative semidefinite, then dCoeffUp_ will be NULL.

Return true if a convexification is there, false otherwise.

Definition at line 41 of file alphaConvexify.cpp.

References bounds_, COUENNE_EPS, COUENNE_INFINITY, DEBUG, eigen_, expression::Index(), info, irow, k, and matrix_.

Referenced by generateCuts().

void exprQuad::quadCuts ( expression w,
OsiCuts &  cs,
const CouenneCutGenerator cg 
)

method exprQuad::quadCuts Based on the information (dIndex_, dCoeffLo_, dCoeffUp_) created/modified by alphaConvexify(), create convexification cuts for this expression.

The original constraint is :

\[ \eta = a_0 + a^T x + x^T Q x \]

where $ \eta $ is the auxiliary corresponding to this expression and $ w_j $ are the auxiliaries corresponding to the other non-linear terms contained in the expression.

The under-estimator of $ x^T Q x$ is given by

\[ x^T Q x + \sum \lambda_{\min,i} (x_i - l_i ) (u_i - x_i ) \]

and its over-estimator is given by

\[ x^T Q x + \sum \lambda_{\max, i} (x_i - l_i ) (u_i - x_i ) \]

(where $ \lambda_{\min, i} = \frac{\lambda_{\min}}{w_i^2} $, $ \lambda_{\max, i} = \frac{\lambda_{\max}}{w_i^2} $, and $w_i = u_i - l_i$), where $\lambda_{\max}$ ($\lambda_{\max}$) is the minimum (maximum) eigenvalue of the matrix $A={\rm Diag}({\bf u} - {\bf l}) Q {\rm Diag}({\bf u} - {\bf l})$, obtained by pre- and post-multiplying $ Q $ by the diagonal matrix whose $i$-th element is $u_i - l_i$.

Let $ \tilde a_0(\lambda)$, $ \tilde a(\lambda) $ and $ \tilde Q (\lambda) $ be

\[ \tilde a_0(\lambda) = a_0 - \sum_{i = 1}^n \lambda_i l_i u_i \]

\[ \tilde a(\lambda) = a + \left[ \begin{array}{c} \lambda_1 (u_1 + l_1) \\ \vdots \\ \lambda_n (u_n + l_n) \end{array} \right], \]

\[ \tilde Q(\lambda) = Q - \left( \begin{array}{ccc} {\lambda_1} & & 0 \\ & \ddots & \\ 0 & & \lambda_n \end{array} \right). \]

The convex relaxation of the initial constraint is then given by the two constraints

\[ \eta \geq \tilde a_0(\lambda_{\min}) + \tilde a(\lambda_{\min})^T x + x^T \tilde Q(\lambda_{\min}) x \]

\[ \eta \leq \tilde a_0(\lambda_{\max}) + \tilde a(\lambda_{\max})^T x + x^T \tilde Q(\lambda_{\max}) x \]

The cut is computed as follow. Let $ (x^*, \eta^*) $ be the solution at hand. The two outer-approximation cuts are:

\[ \eta \geq \tilde a_0(\lambda_{\min}) + \tilde a(\lambda_{\min})^T x + {x^*}^T \tilde Q(\lambda_{\min}) (2x - x^*) \]

and

\[ \eta \leq \tilde a_0(\lambda_{\max}) + \tilde a(\lambda_{\max})^T x + {x^*}^T \tilde Q(\lambda_{\max}) (2x - x^*); \]

grouping coefficients, we get:

\[ {x^*}^T \tilde Q(\lambda_{\min}) x^* - \tilde a_0(\lambda_{\min}) \geq (\tilde a(\lambda_{\min}) + 2 \tilde Q(\lambda_{\min} ) x^*)^T x - \eta \]

and

\[ {x^*}^T \tilde Q(\lambda_{\max}) x^* - \tilde a_0(\lambda_{\max}) \leq (\tilde a(\lambda_{\max}) + 2 \tilde Q (\lambda_{\max}) x^* )^T x - \eta \]

Definition at line 21 of file quadCuts.cpp.

References a, bounds_, exprGroup::c0_, COUENNE_EPS, DEBUG, e, eigen_, expression::Index(), exprGroup::lcoeff_, matrix_, nqterms_, print(), and w.

Referenced by generateCuts().

int exprQuad::compare ( exprQuad  )  [virtual]

Compare two exprQuad.

Definition at line 297 of file exprQuad.cpp.

References exprGroup::compare(), COUENNE_EPS, e, expression::Index(), and matrix_.

virtual enum expr_type exprQuad::code (  )  [inline, virtual]

Code for comparisons.

Reimplemented from exprGroup.

Definition at line 229 of file exprQuad.hpp.

References COU_EXPRQUAD.

int exprQuad::rank (  )  [virtual]

Used in rank-based branching variable choice.

Reimplemented from exprGroup.

Definition at line 343 of file exprQuad.cpp.

References matrix_, r, and exprGroup::rank().

bool exprQuad::isInteger (  )  [virtual]

is this expression integer?

Reimplemented from exprGroup.

Definition at line 397 of file exprQuad.cpp.

References exprGroup::isInteger(), and matrix_.

Referenced by impliedBound().

int exprQuad::DepList ( std::set< int > &  deplist,
enum dig_type  type = ORIG_ONLY 
) [virtual]

fill in the set with all indices of variables appearing in the expression

Reimplemented from exprGroup.

Definition at line 381 of file exprQuad.cpp.

References exprGroup::DepList(), and matrix_.

CouNumber exprQuad::selectBranch ( const CouenneObject obj,
const OsiBranchingInformation *  info,
expression *&  var,
double *&  brpts,
double *&  brDist,
int &  way 
) [virtual]

Set up branching object by evaluating many branching points for each expression's arguments.

Reimplemented from expression.

Definition at line 21 of file branchExprQuad.cpp.

References bounds_, COUENNE_EPS, COUENNE_INFINITY, eigen_, expression::Index(), and TWO_RAND.

void exprQuad::fillDepSet ( std::set< DepNode *, compNode > *  dep,
DepGraph g 
) [virtual]

Fill dependence set of the expression associated with this auxiliary variable.

Reimplemented from exprGroup.

Definition at line 365 of file exprQuad.cpp.

References exprGroup::fillDepSet(), g, expression::Index(), and matrix_.

void exprQuad::replace ( exprVar x,
exprVar w 
) [virtual]

replace variable x with new (aux) w

Reimplemented from exprGroup.

Definition at line 441 of file exprQuad.cpp.

References expression::Index(), matrix_, exprGroup::replace(), w, and x.

void exprQuad::realign ( const CouenneProblem p  )  [virtual]

replace variable x with new (aux) w

Reimplemented from exprGroup.

Definition at line 472 of file exprQuad.cpp.

References AUX, expression::Index(), matrix_, expression::Original(), exprOp::Type(), and VAR.

bool exprQuad::impliedBound ( int  ,
CouNumber ,
CouNumber ,
t_chg_bounds  
) [virtual]

implied bound processing

Reimplemented from exprSum.

Definition at line 36 of file impliedBounds-exprQuad.cpp.

References exprGroup::c0_, t_chg_bounds::CHANGED, computeQuadFiniteBound(), COUENNE_EPS, COUENNE_INFINITY, exprSum::getBounds(), expression::Index(), isInteger(), exprGroup::lcoeff_, li, matrix_, print(), t_chg_bounds::setLower(), t_chg_bounds::setUpper(), and updateBound().

CouNumber exprQuad::computeQBound ( int  sign  ) 

method to compute the bound based on sign: -1 for lower, +1 for upper

Definition at line 16 of file exprBQuad.cpp.

References exprGroup::c0_, COUENNE_INFINITY, expression::Index(), exprGroup::lcoeff_, and matrix_.

void exprQuad::closestFeasible ( expression varind,
expression vardep,
CouNumber left,
CouNumber right 
) const [virtual]

compute $y^{lv}$ and $y^{uv}$ for Violation Transfer algorithm

Reimplemented from expression.

Definition at line 513 of file exprQuad.cpp.

void exprQuad::computeQuadFiniteBound ( CouNumber qMin,
CouNumber qMax,
CouNumber l,
CouNumber u,
int &  indInfLo,
int &  indInfUp 
) [protected]

return lower and upper bound of quadratic expression

Definition at line 29 of file compQuadFinBounds.cpp.

References COUENNE_INFINITY, expression::Index(), exprGroup::lcoeff_, li, matrix_, and updateInf().

Referenced by impliedBound().

virtual bool exprQuad::isCuttable ( CouenneProblem problem,
int  index 
) const [inline, protected, virtual]

can this expression be further linearized or are we on its concave ("bad") side

Reimplemented from expression.

Definition at line 284 of file exprQuad.hpp.


Member Data Documentation

sparseQ exprQuad::matrix_ [mutable, protected]

Definition at line 59 of file exprQuad.hpp.

Referenced by alphaConvexify(), compare(), computeQBound(), computeQuadFiniteBound(), DepList(), differentiate(), exprQuad(), fillDepSet(), getQ(), gradientNorm(), impliedBound(), isInteger(), Linearity(), operator()(), print(), quadCuts(), rank(), realign(), and replace().

std::vector<std::pair <CouNumber, std::vector <std::pair <exprVar *, CouNumber> > > > exprQuad::eigen_ [mutable, protected]

eigenvalues and eigenvectors

Definition at line 71 of file exprQuad.hpp.

Referenced by alphaConvexify(), exprQuad(), quadCuts(), and selectBranch().

std::map<exprVar *, std::pair <CouNumber, CouNumber> > exprQuad::bounds_ [protected]

current bounds (checked before re-computing eigenvalues/vectors)

Definition at line 74 of file exprQuad.hpp.

Referenced by alphaConvexify(), exprQuad(), quadCuts(), and selectBranch().

int exprQuad::nqterms_ [protected]

number of non-zeroes in Q

Definition at line 77 of file exprQuad.hpp.

Referenced by exprQuad(), getnQTerms(), and quadCuts().


The documentation for this class was generated from the following files:
Generated on Thu Aug 5 03:10:52 2010 by  doxygen 1.4.7