Struct LineSearchOptions 

pub struct LineSearchOptions {

    pub alpha_red_factor: Number,
    pub alpha_red_factor_min: Option<Number>,
    pub watchdog_shortened_iter_trigger: Index,
    pub watchdog_trial_iter_max: Index,
    pub soft_resto_pderror_reduction_factor: Number,
    pub max_soft_resto_iters: Index,
    pub accept_every_trial_step: bool,
    pub alpha_for_y: AlphaForY,
    pub accept_after_max_steps: Index,
    pub eta_phi: Number,
    pub delta: Number,
    pub theta_min_fact: Number,
    pub theta_max_fact: Number,
    pub theta_max_row_scale_kappa: Number,
    pub theta_max_adaptive_trigger: u32,
    pub theta_max_adaptive_factor: Number,
    pub theta_max_adaptive_max_raises: u32,
    pub gamma_phi: Number,
    pub gamma_theta: Number,
    pub s_phi: Number,
    pub s_theta: Number,
    pub alpha_min_frac: Number,
    pub obj_max_inc: Number,
    pub max_filter_resets: Index,
    pub filter_reset_trigger: Index,
    pub nu_init: Number,
    pub nu_inc: Number,
    pub rho: Number,
    pub eta_penalty: Number,
    pub max_soc: Index,
    pub kappa_soc: Number,
    pub soc_method: Index,
}

Knobs baked into the assembled BacktrackingLineSearch. Defaults mirror IpBacktrackingLineSearch.cpp:RegisterOptions.

Fields

alpha_red_factor: Number

alpha_red_factor — fractional reduction applied to the trial step size at every backtracking step (alpha *= alpha_red_factor). Mirrors upstream’s IpBacktrackingLineSearch::alpha_red_factor_.

alpha_red_factor_min: Option<Number>

alpha_red_factor_min — floor on one backtracking reduction, which turns the fixed geometric trial sequence into a safeguarded quadratic interpolation (gh#818). See BacktrackingLineSearch::next_alpha.

None — the default — means “let the Hessian mode decide”, and the two modes decide differently:

• limited-memory → 0.05 (interpolation on). The quasi-Newton model’s scale can be wrong by orders of magnitude in any direction its curvature pairs do not span, so the acceptable alpha can be far below 1 and the fixed factor spends log(1/alpha) objective evaluations walking to it.
• exact → off (alpha_red_factor, i.e. upstream’s fixed sequence). A Newton step’s length is meaningful, the acceptable alpha is normally within a few halvings of 1, and the trial sequence is not what costs.

The split is measured, not assumed. Forcing alpha_red_factor_min 0.05 onto the exact path moves 3 of the 156 fixture-legs in scripts/sweep-fixtures.sh, and the one that matters is a status loss: eigena2 goes from SolveSucceeded/27 to SolvedToAcceptableLevel/32. The other two are issue_508_infeasible_gap_1em4 441 → 385 to the same certificate — a gain — and an objective digit on hs13_bigstart. One fixture giving up a solve is the whole argument; a faster infeasibility certificate does not buy it back. Before ALPHA_INTERP_MIN_TRIALS gated the interpolation the same experiment moved 9 legs and cost infeasible_square_scaled_1em4 its infeasibility certificate (InfeasibleProblemDetected/17 → ErrorInStepComputation/12); the gate narrowed the damage, it did not remove the reason for the split.

An explicit alpha_red_factor_min from the user is honoured on both paths — the mode-dependence is only in the default.

watchdog_shortened_iter_trigger: Index

watchdog_trial_iter_max: Index

soft_resto_pderror_reduction_factor: Number

soft_resto_pderror_reduction_factor — required relative reduction in the primal-dual error for a soft-resto step. 0 disables the soft restoration phase.

max_soft_resto_iters: Index

max_soft_resto_iters — cap on consecutive soft-resto iterations before full restoration is forced.

accept_every_trial_step: bool

accept_every_trial_step — short-circuits the filter / alpha loop and accepts the full fraction-to-the-boundary step every outer iteration. Mirrors upstream’s IpBacktrackingLineSearch::accept_every_trial_step_. Drops global convergence guarantees; only safe for problems where the Newton step is already a descent step (LPs, convex QPs). The Mehrotra cascade in application.rs flips this on.

alpha_for_y: AlphaForY

alpha_for_y — policy for the equality-multiplier (y_c / y_d) step length. Upstream default is Primal; the Mehrotra cascade switches to BoundMult.

accept_after_max_steps: Index

accept_after_max_steps — accept a trial point once this many backtracking steps have been taken, even if it fails the acceptor’s tests. -1 (the default) disables it, which is why wiring it moves no default trajectory. Mirrors IpBacktrackingLineSearch.cpp:759-770.

eta_phi: Number

eta_phi — relaxation factor in the Armijo condition (Eqn. (20)).

delta: Number

delta — multiplier on the constraint violation in the filter’s switching rule (Eqn. (19)); maps to FilterLsAcceptor::delta_armijo. Default 1.0, from IpFilterLSAcceptor.cpp:RegisterOptions.

theta_min_fact: Number

theta_min_fact — constraint-violation threshold factor in the switching rule.

theta_max_fact: Number

theta_max_fact — upper-bound factor for constraint violation in the filter (Eqn. (21)).

theta_max_row_scale_kappa: Number

theta_max_row_scale_kappa — multiplier on the constraint-row count used as the floor of the theta_max reference. Opt-in: default 0, which is upstream’s bare max(1, theta_0) floor bit-for-bit. Set to 1 on a large model that stalls from a feasible start. See FilterLsAcceptor::theta_max_row_scale_kappa.

theta_max_adaptive_trigger: u32

theta_max_adaptive_trigger — consecutive line searches whose every trial was refused at the theta_max gate before the ceiling is raised. 0 disables the rule. See FilterLsAcceptor::theta_max_adaptive_trigger (pounce#546).

theta_max_adaptive_factor: Number

Geometric factor applied to theta_max on each adaptive raise. See FilterLsAcceptor::theta_max_adaptive_factor.

theta_max_adaptive_max_raises: u32

Cap on adaptive raises per solve, which is what keeps theta_max finite. See FilterLsAcceptor::theta_max_adaptive_max_raises.

gamma_phi: Number

gamma_phi — filter margin factor for the barrier function (Eqn. (18a)).

gamma_theta: Number

gamma_theta — filter margin factor for the constraint violation (Eqn. (18b)).

s_phi: Number

s_phi — exponent for the linear barrier model in the switching rule (Eqn. (19)).

s_theta: Number

s_theta — exponent for the current constraint violation in the switching rule (Eqn. (19)).

alpha_min_frac: Number

alpha_min_frac — safety factor for the minimal step size before switching to restoration (gamma_alpha, Eqn. (23)).

obj_max_inc: Number

obj_max_inc — max acceptable increase (orders of magnitude) of the barrier objective for a trial point.

max_filter_resets: Index

max_filter_resets — maximum number of filter resets allowed (0 disables the reset heuristic).

filter_reset_trigger: Index

filter_reset_trigger — successive filter-rejected iterations that trigger a filter reset.

nu_init: Number

nu_init — initial value of the penalty parameter ν.

nu_inc: Number

nu_inc — increment added when ν is bumped.

rho: Number

rho — convex-combination weight in the ν update rule.

eta_penalty: Number

eta_penalty — relaxation factor in the Armijo condition on the penalty merit function.

max_soc: Index

max_soc — max second-order-correction trial steps per iteration; 0 disables SOC.

kappa_soc: Number

kappa_soc — sufficient-reduction factor for a SOC step to be continued.

soc_method: Index

soc_method — 0 (paper method) or 1 (alpha-on-rhs variant).

Trait Implementations

impl Clone for LineSearchOptions

fn clone(&self) -> LineSearchOptions

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for LineSearchOptions

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

Formats the value using the given formatter.

impl Default for LineSearchOptions

fn default() -> Self

Returns the “default value” for a type.