Struct IpoptNlpAdapter 

pub struct IpoptNlpAdapter { /* private fields */ }

Implementations

impl IpoptNlpAdapter

pub fn new(nlp: Rc<RefCell<dyn IpoptNlp>>) -> Self

Build the adapter from an IpoptNlp handle. Dimensions are queried directly from Nlp::n(), Nlp::m_eq(), Nlp::m_ineq().

pub fn new_with_declared_bounds(nlp: Rc<RefCell<dyn IpoptNlp>>) -> Self

The same adapter, but reporting the bounds the user declared rather than the live ones the interior method widened by bound_relax_factor.

For the interior iteration the widening is the point — it is what keeps a strictly-interior iterate from being pinned against a bound it must approach. For anything asking where the solution sits relative to the model, it inverts the answer: a point exactly on a declared bound is a full 1e-8 inside the relaxed one, so an activity test against the live bounds calls the binding constraint inactive, and a pivot against them stops short of it. That is the difference between crossover identifying the active set and crossover identifying nothing.

Falls back to the live bounds for any block the NLP does not track (the trait accessors default to None).

Trait Implementations

impl SqpProblemSpec for IpoptNlpAdapter

fn n(&self) -> usize

fn m(&self) -> usize

fn x_init(&self) -> Vec<Number>

fn variable_bounds(&self) -> (Vec<Number>, Vec<Number>)

fn constraint_bounds(&self) -> (Vec<Number>, Vec<Number>)

fn eval_f(&mut self, x: &[Number]) -> Number

fn eval_grad_f(&mut self, x: &[Number]) -> Vec<Number>

fn eval_c(&mut self, x: &[Number]) -> Vec<Number>

c(x) — combined constraint values (length m). The constraint bounds from constraint_bounds apply directly: row i is a strict equality if bl[i] == bu[i], an inequality otherwise.

fn eval_jac_c(&mut self, x: &[Number]) -> Triplet

∇c(x) as a sparse m × n triplet (1-based indices).

fn eval_hess_lag(&mut self, x: &[Number], lambda_g: &[Number]) -> Triplet

∇²L(x, λ_g) = ∇²f(x) + Σ λ_g_i · ∇²c_i(x) as a sparse symmetric n × n triplet.