// Test KKT conditions on the solution.
// The function returns true if the tested KKT conditions are satisfied
// and false otherwise.
// The function assumes that the model currently extracted to CPLEX is fully
// described by obj, vars and rngs.
static bool
checkkkt (IloCplex& cplex, IloObjective const& obj, IloNumVarArray const& vars,
IloRangeArray const& rngs, IloIntArray const& cone, double tol)
{
IloEnv env = cplex.getEnv();
IloModel model = cplex.getModel();
IloNumArray x(env), dslack(env);
IloNumArray pi(env, rngs.getSize()), slack(env);
// Read primal and dual solution information.
cplex.getValues(x, vars);
cplex.getSlacks(slack, rngs);
// pi for second order cone constraints.
getsocpconstrmultipliers(cplex, vars, rngs, pi, dslack);
// pi for linear constraints.
for (IloInt i = 0; i < rngs.getSize(); ++i) {
IloRange r = rngs[i];
if ( !r.getQuadIterator().ok() )
pi[idx(r)] = cplex.getDual(r);
}
// Print out the data we just fetched.
streamsize oprec = env.out().precision(3);
ios_base::fmtflags oflags = env.out().setf(ios::fixed | ios::showpos);
env.out() << "x = [";
for (IloInt i = 0; i < x.getSize(); ++i)
env.out() << " " << x[i];
env.out() << " ]" << endl;
env.out() << "dslack = [";
for (IloInt i = 0; i < dslack.getSize(); ++i)
env.out() << " " << dslack[i];
env.out() << " ]" << endl;
env.out() << "pi = [";
for (IloInt i = 0; i < rngs.getSize(); ++i)
env.out() << " " << pi[i];
env.out() << " ]" << endl;
env.out() << "slack = [";
for (IloInt i = 0; i < rngs.getSize(); ++i)
env.out() << " " << slack[i];
env.out() << " ]" << endl;
env.out().precision(oprec);
env.out().flags(oflags);
// Test primal feasibility.
// This example illustrates the use of dual vectors returned by CPLEX
// to verify dual feasibility, so we do not test primal feasibility
// here.
// Test dual feasibility.
// We must have
// - for all <= constraints the respective pi value is non-negative,
// - for all >= constraints the respective pi value is non-positive,
// - the dslack value for all non-cone variables must be non-negative.
// Note that we do not support ranged constraints here.
for (IloInt i = 0; i < vars.getSize(); ++i) {
IloNumVar v = vars[i];
if ( cone[i] == NOT_IN_CONE && dslack[i] < -tol ) {
env.error() << "Dual multiplier for " << v << " is not feasible: "
<< dslack[i] << endl;
return false;
}
}
for (IloInt i = 0; i < rngs.getSize(); ++i) {
IloRange r = rngs[i];
if ( fabs (r.getLB() - r.getUB()) <= tol ) {
// Nothing to check for equality constraints.
}
else if ( r.getLB() > -IloInfinity && pi[i] > tol ) {
env.error() << "Dual multiplier " << pi[i] << " for >= constraint"
<< endl << r << endl
<< "not feasible"
<< endl;
return false;
}
else if ( r.getUB() < IloInfinity && pi[i] < -tol ) {
env.error() << "Dual multiplier " << pi[i] << " for <= constraint"
<< endl << r << endl
<< "not feasible"
<< endl;
return false;
}
}
// Test complementary slackness.
// For each constraint either the constraint must have zero slack or
// the dual multiplier for the constraint must be 0. We must also
// consider the special case in which a variable is not explicitly
// contained in a second order cone constraint.
for (IloInt i = 0; i < vars.getSize(); ++i) {
if ( cone[i] == NOT_IN_CONE ) {
if ( fabs(x[i]) > tol && dslack[i] > tol ) {
env.error() << "Invalid complementary slackness for " << vars[i]
<< ":" << endl
<< " " << x[i] << " and " << dslack[i]
<< endl;
return false;
}
}
}
for (IloInt i = 0; i < rngs.getSize(); ++i) {
if ( fabs(slack[i]) > tol && fabs(pi[i]) > tol ) {
env.error() << "Invalid complementary slackness for "
<< endl << rngs[i] << ":" << endl
<< " " << slack[i] << " and " << pi[i]
<< endl;
return false;
}
}
// Test stationarity.
// We must have
// c - g[i]'(X)*pi[i] = 0
// where c is the objective function, g[i] is the i-th constraint of the
// problem, g[i]'(x) is the derivate of g[i] with respect to x and X is the
// optimal solution.
// We need to distinguish the following cases:
// - linear constraints g(x) = ax - b. The derivative of such a
// constraint is g'(x) = a.
// - second order constraints g(x[1],...,x[n]) = -x[1] + |(x[2],...,x[n])|
// the derivative of such a constraint is
// g'(x) = (-1, x[2]/|(x[2],...,x[n])|, ..., x[n]/|(x[2],...,x[n])|
// (here |.| denotes the Euclidean norm).
// - bound constraints g(x) = -x for variables that are not explicitly
// contained in any second order cone constraint. The derivative for
// such a constraint is g'(x) = -1.
// Note that it may happen that the derivative of a second order cone
// constraint is not defined at the optimal solution X (this happens if
// X=0). In this case we just skip the stationarity test.
IloNumArray sum(env, vars.getSize());
for (IloExpr::LinearIterator it = obj.getLinearIterator(); it.ok(); ++it)
sum[idx(it.getVar())] = it.getCoef();
for (IloInt i = 0; i < vars.getSize(); ++i) {
IloNumVar v = vars[i];
if ( cone[i] == NOT_IN_CONE )
sum[i] -= dslack[i];
}
for (IloInt i = 0; i < rngs.getSize(); ++i) {
IloRange r = rngs[i];
if ( r.getQuadIterator().ok() ) {
// Quadratic (second order cone) constraint.
IloNum norm = 0.0;
for (IloExpr::QuadIterator q = r.getQuadIterator(); q.ok(); ++q) {
if ( q.getCoef() > 0 )
norm += x[idx(q.getVar1())] * x[idx(q.getVar1())];
}
norm = sqrt(norm);
if ( fabs(norm) <= tol ) {
// Derivative is not defined. Skip test.
env.warning() << "Cannot test stationarity at non-differentiable point."
<< endl;
return true;
}
else {
for (IloExpr::QuadIterator q = r.getQuadIterator(); q.ok(); ++q) {
if ( q.getCoef() < 0 )
sum[idx(q.getVar1())] -= pi[i];
else
sum[idx(q.getVar1())] += pi[i] * x[idx(q.getVar1())] / norm;
}
}
}
else {
// Linear constraint.
for (IloExpr::LinearIterator l = r.getLinearIterator(); l.ok(); ++l)
sum[idx(l.getVar())] -= pi[i] * l.getCoef();
}
}
// Now test that all elements in sum[] are 0.
for (IloInt i = 0; i < vars.getSize(); ++i) {
if ( fabs(sum[i]) > tol ) {
env.error() << "Invalid stationarity " << sum[i] << " for "
<< vars[i] << endl;
return false;
}
}
return true;
}
int
main (int argc, char **argv)
{
int result = 0;
IloEnv env;
try {
static int indices[] = { 0, 1, 2, 3, 4, 5, 6 };
IloModel model(env);
/* ***************************************************************** *
* *
* S E T U P P R O B L E M *
* *
* The model we setup here is *
* Minimize *
* obj: 3x1 - x2 + 3x3 + 2x4 + x5 + 2x6 + 4x7 *
* Subject To *
* c1: x1 + x2 = 4 *
* c2: x1 + x3 >= 3 *
* c3: x6 + x7 <= 5 *
* c4: -x1 + x7 >= -2 *
* q1: [ -x1^2 + x2^2 ] <= 0 *
* q2: [ 4.25x3^2 -2x3*x4 + 4.25x4^2 - 2x4*x5 + 4x5^2 ] + 2 x1 <= 9.0
* q3: [ x6^2 - x7^2 ] >= 4 *
* Bounds *
* 0 <= x1 <= 3 *
* x2 Free *
* 0 <= x3 <= 0.5 *
* x4 Free *
* x5 Free *
* x7 Free *
* End *
* *
* ***************************************************************** */
IloNumVarArray x(env);
x.add(IloNumVar(env, 0, 3, "x1"));
x.add(IloNumVar(env, -IloInfinity, IloInfinity, "x2"));
x.add(IloNumVar(env, 0, 0.5, "x3"));
x.add(IloNumVar(env, -IloInfinity, IloInfinity, "x4"));
x.add(IloNumVar(env, -IloInfinity, IloInfinity, "x5"));
x.add(IloNumVar(env, 0, IloInfinity, "x6"));
x.add(IloNumVar(env, -IloInfinity, IloInfinity, "x7"));
for (IloInt i = 0; i < x.getSize(); ++i)
x[i].setObject(&indices[i]);
IloObjective obj = IloMinimize(env,
3*x[0] - x[1] + 3*x[2] + 2*x[3] +
x[4] + 2*x[5] + 4*x[6], "obj");
model.add(obj);
IloRangeArray linear(env);
linear.add(IloRange(env, 4.0, x[0] + x[1], 4.0, "c1"));
linear.add(IloRange(env, 3.0, x[0] + x[2], IloInfinity, "c2"));
linear.add(IloRange(env, -IloInfinity, x[5] + x[6], 5.0, "c3"));
linear.add(IloRange(env, -2.0, -x[0] + x[6], IloInfinity, "c4"));
for (IloInt i = 0; i < linear.getSize(); ++i)
linear[i].setObject(&indices[i]);
model.add(linear);
IloRangeArray quad(env);
quad.add(IloRange(env, -IloInfinity, -x[0]*x[0] + x[1] * x[1], 0, "q1"));
quad.add(IloRange(env, -IloInfinity,
4.25*x[2]*x[2] - 2*x[2]*x[3] + 4.25*x[3]*x[3] +
-2*x[3]*x[4] + 4*x[4]*x[4] + 2*x[0],
9.0, "q2"));
quad.add(IloRange(env, 4.0, x[5]*x[5] - x[6]*x[6], IloInfinity, "q3"));
for (IloInt i = 0; i < quad.getSize(); ++i)
quad[i].setObject(&indices[i]);
model.add(quad);
/* ***************************************************************** *
* *
* O P T I M I Z E P R O B L E M *
* *
* ***************************************************************** */
IloCplex cplex(model);
cplex.setParam(IloCplex::Param::Barrier::QCPConvergeTol, 1e-10);
cplex.solve();
/* ***************************************************************** *
* *
* Q U E R Y S O L U T I O N *
* *
* ***************************************************************** */
IloNumArray xval(env);
IloNumArray slack(env);
IloNumArray qslack(env);
IloNumArray cpi(env);
IloNumArray rpi(env);
IloNumArray qpi;
cplex.getValues(x, xval);
cplex.getSlacks(slack, linear);
cplex.getSlacks(qslack, quad);
cplex.getReducedCosts(cpi, x);
cplex.getDuals(rpi, linear);
qpi = getqconstrmultipliers(cplex, xval, quad);
/* ***************************************************************** *
* *
* C H E C K K K T C O N D I T I O N S *
* *
* Here we verify that the optimal solution computed by CPLEX *
* (and the qpi[] values computed above) satisfy the KKT *
* conditions. *
* *
* ***************************************************************** */
// Primal feasibility: This example is about duals so we skip this test.
// Dual feasibility: We must verify
// - for <= constraints (linear or quadratic) the dual
// multiplier is non-positive.
// - for >= constraints (linear or quadratic) the dual
// multiplier is non-negative.
for (IloInt i = 0; i < linear.getSize(); ++i) {
if ( linear[i].getLB() <= -IloInfinity ) {
// <= constraint
if ( rpi[i] > ZEROTOL ) {
cerr << "Dual feasibility test failed for <= row " << i
<< ": " << rpi[i] << endl;
result = -1;
}
}
else if ( linear[i].getUB() >= IloInfinity ) {
// >= constraint
if ( rpi[i] < -ZEROTOL ) {
cerr << "Dual feasibility test failed for >= row " << i
<< ":" << rpi[i] << endl;
result = -1;
}
}
else {
// nothing to do for equality constraints
}
}
for (IloInt q = 0; q < quad.getSize(); ++q) {
if ( quad[q].getLB() <= -IloInfinity ) {
// <= constraint
if ( qpi[q] > ZEROTOL ) {
cerr << "Dual feasibility test failed for <= quad " << q
<< ": " << qpi[q] << endl;
result = -1;
}
}
else if ( quad[q].getUB() >= IloInfinity ) {
// >= constraint
if ( qpi[q] < -ZEROTOL ) {
cerr << "Dual feasibility test failed for >= quad " << q
<< ":" << qpi[q] << endl;
result = -1;
}
}
else {
// nothing to do for equality constraints
}
}
// Complementary slackness.
// For any constraint the product of primal slack and dual multiplier
// must be 0.
for (IloInt i = 0; i < linear.getSize(); ++i) {
if ( fabs (linear[i].getUB() - linear[i].getLB()) > ZEROTOL &&
fabs (slack[i] * rpi[i]) > ZEROTOL )
{
cerr << "Complementary slackness test failed for row " << i
<< ": " << fabs (slack[i] * rpi[i]) << endl;
result = -1;
}
}
for (IloInt q = 0; q < quad.getSize(); ++q) {
if ( fabs (quad[q].getUB() - quad[q].getLB()) > ZEROTOL &&
fabs (qslack[q] * qpi[q]) > ZEROTOL )
{
cerr << "Complementary slackness test failed for quad " << q
<< ":" << fabs (qslack[q] * qpi[q]) << endl;
result = -1;
}
}
for (IloInt j = 0; j < x.getSize(); ++j) {
if ( x[j].getUB() < IloInfinity ) {
double const slk = x[j].getUB() - xval[j];
double const dual = cpi[j] < -ZEROTOL ? cpi[j] : 0.0;
if ( fabs (slk * dual) > ZEROTOL ) {
cerr << "Complementary slackness test failed for ub " << j
<< ": " << fabs (slk * dual) << endl;
result = -1;
}
}
if ( x[j].getLB() > -IloInfinity ) {
double const slk = xval[j] - x[j].getLB();
double const dual = cpi[j] > ZEROTOL ? cpi[j] : 0.0;
if ( fabs (slk * dual) > ZEROTOL ) {
cerr << "Complementary slackness test failed for lb " << j
<< ": " << fabs (slk * dual) << endl;
result = -1;
}
}
}
// Stationarity.
// The difference between objective function and gradient at optimal
// solution multiplied by dual multipliers must be 0, i.e., for the
// optimal solution x
// 0 == c
// - sum(r in rows) r'(x)*rpi[r]
// - sum(q in quads) q'(x)*qpi[q]
// - sum(c in cols) b'(x)*cpi[c]
// where r' and q' are the derivatives of a row or quadratic constraint,
// x is the optimal solution and rpi[r] and qpi[q] are the dual
// multipliers for row r and quadratic constraint q.
// b' is the derivative of a bound constraint and cpi[c] the dual bound
// multiplier for column c.
IloNumArray kktsum(env, x.getSize());
// Objective function.
for (IloExpr::LinearIterator it = obj.getLinearIterator(); it.ok(); ++it)
kktsum[idx(it.getVar())] = it.getCoef();
// Linear constraints.
// The derivative of a linear constraint ax - b (<)= 0 is just a.
for (IloInt i = 0; i < linear.getSize(); ++i) {
for (IloExpr::LinearIterator it = linear[i].getLinearIterator();
it.ok(); ++it)
kktsum[idx(it.getVar())] -= rpi[i] * it.getCoef();
}
// Quadratic constraints.
// The derivative of a constraint xQx + ax - b <= 0 is
// Qx + Q'x + a.
for (IloInt q = 0; q < quad.getSize(); ++q) {
for (IloExpr::LinearIterator it = quad[q].getLinearIterator();
it.ok(); ++it)
kktsum[idx(it.getVar())] -= qpi[q] * it.getCoef();
for (IloExpr::QuadIterator it = quad[q].getQuadIterator();
it.ok(); ++it)
{
kktsum[idx(it.getVar1())] -= qpi[q] * xval[idx(it.getVar2())] * it.getCoef();
kktsum[idx(it.getVar2())] -= qpi[q] * xval[idx(it.getVar1())] * it.getCoef();
}
}
// Bounds.
// The derivative for lower bounds is -1 and that for upper bounds
// is 1.
// CPLEX already returns dj with the appropriate sign so there is
// no need to distinguish between different bound types here.
for (IloInt j = 0; j < x.getSize(); ++j)
kktsum[j] -= cpi[j];
for (IloInt j = 0; j < kktsum.getSize(); ++j) {
if ( fabs (kktsum[j]) > ZEROTOL ) {
cerr << "Stationarity test failed at index " << j
<< ": " << kktsum[j] << endl;
result = -1;
}
}
if ( result == 0) {
// KKT conditions satisfied. Dump out the optimal solutions and
// the dual values.
streamsize oprec = cout.precision(3);
ios_base::fmtflags oflags = cout.setf(ios::fixed | ios::showpos);
cout << "Optimal solution satisfies KKT conditions." << endl;
cout << " x[] = " << xval << endl;
cout << " cpi[] = " << cpi << endl;
cout << " rpi[] = " << rpi << endl;
cout << " qpi[] = " << qpi << endl;
cout.precision(oprec);
cout.flags(oflags);
}
}
catch (IloException& e) {
cerr << "Concert exception caught: " << e << endl;
result = -1;
}
catch (...) {
cerr << "Unknown exception caught" << endl;
result = -1;
}
env.end();
return result;
} // END main