Styxis
Company
Knowledge

Check original MILP constraints before interpreting optimal status

Separate feasibility from optimality in that order. Do not use equality to one solution vector as the only reference.

Constraints beyond success status

Separate success status from feasibility. Substitute the returned vector into the original equalities, bounds, and integrality conditions before interpreting optimality. An attractive objective cannot justify an infeasible allocation, and floating-point representations still need an appropriate tolerance policy. Recompute the left-hand sides supplied to Truthound externally by substituting the returned solution into the original coefficients. [1, 2, 3]

What the HiGHS follow-up establishes

SciPy #24141 reports milp success with a vector violating the original equality constraints. Follow-up testing reports a feasible result in the HiGHS 1.12.0 upgrade PR. Maintainers emphasize checking constraints rather than matching one vector where multiple optima are possible; this article does not independently verify a released build. [1]

Recompute with original coefficients

Preserve original coefficients, right-hand sides, and integer conditions, recomputing every left-hand side from the returned x. Hold infeasible schedules or allocations while checking settings and independent paths. Operators join recomputed and required values by solution and constraint ID, preserving failures separately. [1, 3]

A small x1+x2=5 check

The illustrative equality is x1+x2=5. Integer candidate [2,3] gives left-hand side 5 and passes against the input right-hand side 5. Candidate [2,4] gives 6 and fails. The exact comparison in this small integer fixture is not a universal zero-tolerance requirement for real solvers. [1, 3]

Constraint-table comparison versus solving

ColumnPairValidator in Truthound Python SDK 3.3.2 compares externally recomputed_lhs with input required_rhs values. Operators also provide the tolerance policy, integrality checks, and remaining constraints. Core neither solves MILP nor repairs a solution; one equality cannot establish all feasibility conditions or optimality. [1, 2, 3]

Keep tolerances when comparing alternatives

Check all original constraints while comparing verified solvers, costs, and tolerances. Agreement between solvers alone does not prove optimality. Existing independent feasibility checks can make another audit redundant; a different feasible optimal vector need not be rejected merely because it differs from a reference vector. [1, 3]

Application conditions

Externally recheck the returned solution using original coefficients, right-hand sides, bounds, integrality, and prior tolerances.

Alternatives

Compare other solvers against every original constraint under the same tolerance policy.

Limitations

  • One passing equality does not establish all feasibility conditions, optimality, or uniqueness.

Evidence

Event checked
Subject checked
Modified