Struct BoundRelax 

pub struct BoundRelax {
    pub factor: f64,
    pub cap: f64,
}

Ipopt’s bound_relax_factor widening, as the convex extractors apply it.

The NLP path widens x_L/x_U and the inequality-row bounds d_L/d_U before the algorithm ever sees them (OrigIpoptNlp::relax_bounds, driven from Application with bound_relax_factor — Ipopt default 1e-8 — capped by constr_viol_tol, default 1e-4). The convex path did not, so the same binary solved a materially different model depending on solver_selection.

That is not a hairline difference on a constraint-degenerate model. On LISWET1 (gh #744) every one of the 10 000 monotonicity rows is active at the optimum and the multipliers sum to 1.6e9, so a 1e-8 widening of the rows buys 9.0 of objective — the convex arm returned the exact optimum 36.1224 and the NLP arm (and Ipopt-MA57) the relaxed one, 27.1221, and the 33% gap was read as a convex-solver bug. Both arms now relax, so both report 27.1221, and bound_relax_factor=0 gets 36.1224 from either.

Faithful to relax_bounds in three details that matter:

• Equality rows are not relaxed. Upstream they live in c(x) = 0, which relax_bounds never touches; only d_L/d_U (inequality rows) and the variable box are widened.
• Rows use the scale-relative width min(factor, cap)·|b| (with |b| read as 1 at a declared-zero bound), the gh #385 form. The variable box keeps the upstream absolute formula min(factor·max(|b|,1), cap).
• Fixed variables (x_l == x_u) keep their bounds. Under the default fixed_variable_treatment=make_parameter upstream removes them before relax_bounds runs, so they are never widened.

Fields

factor: f64

bound_relax_factor. Non-positive disables the widening entirely.

cap: f64

constr_viol_tol — the cap on the widening.

Implementations

impl BoundRelax

pub const NONE: Self

No widening — the model exactly as declared. What the convex path did unconditionally before gh #744, and what bound_relax_factor=0 selects.

pub fn active(self) -> bool

Whether any widening is actually applied. Public so a caller can tell “the declared model and the solved model coincide” from “they differ”, which struct equality against Self::NONE cannot: bound_relax_factor=0 zeroes the factor but leaves cap at constr_viol_tol, so the pair is inactive without being NONE.

Trait Implementations

impl Clone for BoundRelax

fn clone(&self) -> BoundRelax

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Copy for BoundRelax

impl Debug for BoundRelax

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl PartialEq for BoundRelax

fn eq(&self, other: &BoundRelax) -> bool

Equality operator ==.

fn ne(&self, other: &Rhs) -> bool

Inequality operator !=.

impl StructuralPartialEq for BoundRelax