Presolver incorrectly marks feasible Quadratic constraints as INFEASIBLE due to logic in doCaseN

[leorm2002]

Hello,

I encountered an issue where a feasible model containing quadratic expressions (with both quadratic and linear terms) is incorrectly marked as INFEASIBLE during the presolve phase. If I disable the presolver, the model is solved correctly.

Upon inspecting the source code, I believe the issue lies within the doCaseN method and its helper methods, specifically isNegativeOn and isPositiveOn.

Currently, the logic seems to be:

boolean isNegativeOn(final Set<IntIndex> subset) {
    if (!this.isAnyQuadraticFactorNonZero()) {
        // ... logic checking linear factors ...
    }
    // If there ARE quadratic factors, the loop is skipped and it returns TRUE
    return true;
}

If an expression contains quadratic factors (e.g., x2), isAnyQuadraticFactorNonZero() returns true. Consequently, the method skips the linear checks and falls through to return true.

Here is a standalone test case demonstrating the issue. It solves a simple quadratic problem twice: first with the default presolvers (Result: INFEASIBLE), and then without them (Result: OPTIMAL).

package org.ojalgo.optimisation;

import java.math.BigDecimal;
import org.ojalgo.optimisation.solver.cplex.SolverCPLEX;


public class OJPresolveTest {

    public static void main(String[] args) {
        // Run with presolver
        System.out.println("--- Run WITH Presolver ---");
        caseN();
    }

    static void caseN() {
        ExpressionsBasedModel model = getModel();

        // 1. Solve with default settings (Presolver enabled)
        Optimisation.Result result = model.minimise();
        System.out.println("State with presolver: " + result.getState());

        // 2. Clear presolvers and solve again
        ExpressionsBasedModel.clearPresolvers();
        // Re-create model to ensure clean state
        model = getModel(); 
        
        Optimisation.Result resultNoPre = model.minimise();
        System.out.println("State without presolver: " + resultNoPre.getState());
    }

    private static ExpressionsBasedModel getModel() {
        ExpressionsBasedModel model = new ExpressionsBasedModel();
        Variable x1 = model.addVariable("x1");
        Variable x2 = model.addVariable("x2");
        Variable x3 = model.addVariable("x3");

        x1.upper(1).lower(0);
        x2.upper(1).lower(0);
        x3.upper(1).lower(0);

        // Create the quadratic expression
        Expression quadExpr = model.addExpression("QuadraticExpression");
        quadExpr.set(x1, x1, 0.1);
        quadExpr.set(x2, x2, 0.1);
        quadExpr.set(x3, x3, 0.1);
        quadExpr.set(x1, -0.0649319296487038);
        quadExpr.set(x2, -0.035049269354063314);
        quadExpr.set(x3, -0.054569655903981634);

        // This constraint causes the issue during presolve
        quadExpr.upper(new BigDecimal("-0.02"));
        
        return model;
    }
}

[apete]

Thanks for narrowing down the problem.

As far as I can tell the default behaviour of isNegativeOn and isPositiveOn when there are quadratic terms is wrong, and fixing that solves your problem, right?

I'll fix this, and there will be a new release soon.

[leorm2002]

yes i think that fixing the behaviour for quadratic terms in isNegativeOn and isPositiveOn should fix everything.

Thanks

[apete]

There is a fix in the develop branch now if you want try it.