pounce-algorithm 0.11.0

Algorithm-side core for POUNCE (port of Ipopt's src/Algorithm/): IteratesVector, IpoptData, CalculatedQuantities, KKT solvers, line search, mu update, conv check, initializer, IpoptAlg main loop, AlgBuilder.
Documentation
//! `ConvCheck` trait — port of `IpConvCheck.hpp`.
//!
//! Upstream `ConvergenceCheck::CheckConvergence` reads the NLP error
//! and iter count off `IpData()`/`IpCq()`. The default trait method
//! pushes that read to the caller (the main loop in `ipopt_alg.rs`)
//! so simple convergence policies stay pure scalar state machines
//! over `(nlp_err, iter_count)`. The richer
//! [`ConvCheck::check_convergence_with_state`] entry point exposes
//! the live `(IpoptData, IpoptCq)` so policies that need iterate
//! components — notably the restoration-side
//! `RestoFilterConvergenceCheck::TestOrigProgress` — can read them.
//! Default impl just delegates to the scalar method, preserving
//! backwards compatibility for every existing impl.

use crate::ipopt_cq::IpoptCqHandle;
use crate::ipopt_data::IpoptDataHandle;
use pounce_common::types::{Index, Number};

#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum ConvergenceStatus {
    Continue,
    Converged,
    /// Converged to the looser `acceptable_*` tolerance band rather
    /// than the tight `tol` — upstream `CONVERGED_TO_ACCEPTABLE_POINT`.
    /// Maps to `SolverReturn::StopAtAcceptablePoint` →
    /// `ApplicationReturnStatus::SolvedToAcceptableLevel`.
    ConvergedToAcceptable,
    MaxIterExceeded,
    /// `max_cpu_time` budget reached. Maps to
    /// `SolverReturn::CpuTimeExceeded` → `MaximumCpuTimeExceeded`.
    CpuTimeExceeded,
    /// `max_wall_time` budget reached. Maps to
    /// `SolverReturn::WallTimeExceeded` → `MaximumWallTimeExceeded`.
    WallTimeExceeded,
    /// Rapid infeasibility detection fired — the iterate is
    /// converging to a stationary point of the constraint violation
    /// with the violation bounded away from zero. Maps to
    /// `SolverReturn::LocalInfeasibility`.
    LocallyInfeasible,
    Failed,
}

pub trait ConvCheck {
    fn check_convergence(&mut self, nlp_err: Number, iter_count: Index) -> ConvergenceStatus;

    /// State-aware convergence check. The main loop calls this on
    /// every iteration so policies that need access to the iterate
    /// (e.g. `RestoConvCheckAdapter`'s orig-NLP `inf_pr` evaluation
    /// for the kappa-reduction early-exit) can read `data.curr` and
    /// the cq layer. Default impl delegates to
    /// [`Self::check_convergence`], so scalar-only policies don't
    /// need to override.
    /// Whether this policy ever refused a termination certificate it judged
    /// masked by an extreme objective scale (gh #200).
    ///
    /// The main loop reads it on the failure exits: a run that was held back
    /// from stopping must not end up *worse off* than it would have been, so
    /// when it later stalls the stored acceptable point is restored rather
    /// than a bare failure surfaced. Policies that never veto keep the
    /// default `false` and the failure paths behave exactly as before.
    fn certificate_vetoed(&self) -> bool {
        false
    }

    /// Whether this policy ever refused an *acceptable-level* termination it
    /// judged masked (gh #200). Undone differently from a strict refusal — see
    /// `OptErrorConvCheck::acceptable_veto_fired`.
    fn acceptable_certificate_vetoed(&self) -> bool {
        false
    }

    fn check_convergence_with_state(
        &mut self,
        nlp_err: Number,
        iter_count: Index,
        _data: &IpoptDataHandle,
        _cq: &IpoptCqHandle,
    ) -> ConvergenceStatus {
        self.check_convergence(nlp_err, iter_count)
    }

    /// Ask [`Self::check_convergence_with_state`] what it *would* say,
    /// without letting the answer count as an iteration.
    ///
    /// `IpoptAlgorithm::ComputeFeasibilityMultipliers` calls the
    /// convergence check twice on the same iterate — once to decide
    /// whether re-estimating the multipliers is worth attempting
    /// (`IpIpoptAlg.cpp:880`), once to decide whether to keep the result
    /// (`cpp:924`) — on top of the main loop's own call. Upstream wears
    /// the double-count because the only state it advances is
    /// `acceptable_counter_`. pounce's check carries considerably more:
    /// the rapid-infeasibility streak, the gh#200 veto budget, the gh#533
    /// acceptable-progress window. Advancing those three times per
    /// iteration is not upstream behaviour under a different name — it
    /// changes when detectors fire. Measured on the gh#508 probe: the
    /// rapid detector convicted at iteration 33 instead of 86, purely
    /// from the extra increments.
    ///
    /// Default delegates, since a policy with no cross-iteration state
    /// has nothing to protect.
    fn probe_convergence(
        &mut self,
        nlp_err: Number,
        iter_count: Index,
        data: &IpoptDataHandle,
        cq: &IpoptCqHandle,
    ) -> ConvergenceStatus {
        self.check_convergence_with_state(nlp_err, iter_count, data, cq)
    }

    /// Whether the current iterate passes the *strict* per-component convergence
    /// tolerances — the [`Self::check_convergence_with_state`] strict test with
    /// the masked-certificate veto (gh #200) removed. In other words: would this
    /// iterate have certified `Success` were it not for the objective-scale
    /// masking?
    ///
    /// gh #327 reads it on the veto's fallback path to distinguish a point the
    /// veto refused *only because of masking* — a would-be strict certificate,
    /// e.g. the true optimum a continued run reached but could never certify
    /// there — from one that never converged at all (an unbounded / diverging
    /// iterate whose gradient never vanishes). Only the former may displace the
    /// point the baseline stopped at. Default `false` for policies that expose no
    /// strict test.
    fn current_passes_strict(
        &self,
        _nlp_err: Number,
        _data: &IpoptDataHandle,
        _cq: &IpoptCqHandle,
    ) -> bool {
        false
    }

    /// Whether the supplied `nlp_err` is at or below the acceptable
    /// tolerance — port of upstream
    /// `OptimalityErrorConvergenceCheck::CurrentIsAcceptable`. Used by
    /// the main loop to gate `StoreAcceptablePoint` /
    /// `RestoreAcceptablePoint`. Default returns `false` so policies
    /// that don't track an acceptable level (e.g. resto-of-resto inner
    /// adapters) silently skip the rollback machinery.
    fn current_is_acceptable(&self, _nlp_err: Number) -> bool {
        false
    }

    /// State-aware acceptance check. Mirrors upstream
    /// `OptimalityErrorConvergenceCheck::CurrentIsAcceptable` which
    /// reads the per-component residuals and current `f` to gate the
    /// `acceptable_dual_inf_tol` / `acceptable_constr_viol_tol` /
    /// `acceptable_compl_inf_tol` / `acceptable_obj_change_tol`
    /// triplet. Default delegates to the scalar [`Self::current_is_acceptable`].
    fn current_is_acceptable_with_state(
        &self,
        nlp_err: Number,
        _data: &IpoptDataHandle,
        _cq: &IpoptCqHandle,
    ) -> bool {
        self.current_is_acceptable(nlp_err)
    }

    /// Record the current objective at the iterate the main loop just
    /// stashed as the latest "acceptable point" — mirrors upstream
    /// `OptimalityErrorConvergenceCheck::SetCurrAcceptableF`. The
    /// recorded value feeds the `acceptable_obj_change_tol` stability
    /// cross-check on subsequent iterates. Default no-op for policies
    /// that don't track acceptable points.
    fn set_curr_acceptable_obj(&mut self, _obj: Number) {}

    /// Outer NLP convergence tolerance, as used by the main loop's
    /// almost-feasible bypass guard (port of
    /// `IpBacktrackingLineSearch.cpp:580`). Default `1e-8` matches
    /// upstream's default `tol`.
    fn tol_or_default(&self) -> Number {
        1e-8
    }

    /// Primal-feasibility tolerance `constr_viol_tol`, in the unscaled max-norm
    /// space `curr_unscaled_primal_infeasibility_max` reports — the option that
    /// declares what a *violated* constraint is.
    ///
    /// Read by the status-decision sites that have to answer "is this violation
    /// real" (gh #508). That question is about the constraint violation, so it
    /// must be asked with the constraint-violation tolerance; asking it with
    /// `tol` — a tolerance on the **KKT error**, a different quantity in
    /// different units — makes the answer move when the user retunes
    /// convergence and stand still when they retune feasibility. Default `1e-4`
    /// matches upstream's default `constr_viol_tol`; policies that track no such
    /// tolerance keep it.
    fn constr_viol_tol_or_default(&self) -> Number {
        1e-4
    }

    /// Acceptable-level primal-feasibility band `acceptable_constr_viol_tol`, in
    /// the unscaled max-norm space `curr_unscaled_primal_infeasibility_max`
    /// reports. Read by the best-acceptable fallback's feasibility-aware ranking
    /// (gh #267), which caps it at the upstream default so a user-widened band
    /// cannot let the fallback spend feasibility to buy objective. Default
    /// `1e-2` matches upstream's default `acceptable_constr_viol_tol`; policies
    /// that track no such tolerance keep it.
    fn acceptable_constr_viol_tol_or_default(&self) -> Number {
        1e-2
    }

    /// Live-update a named convergence tolerance mid-solve, for the
    /// debugger's in-place option hot-swap. Returns `true` if `name`
    /// matched a tolerance this policy owns (so the caller can report
    /// whether it took). Default: this policy exposes no live
    /// tolerances → `false`.
    fn set_tolerance(&mut self, _name: &str, _value: Number) -> bool {
        false
    }
}