Keyboard shortcuts

Press or to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

Solver Options

POUNCE accepts options the same way upstream Ipopt does. Option names and semantics follow Ipopt’s, so an existing Ipopt options file or KEY=VALUE invocation works unchanged.

Setting options

On the command line — append KEY=VALUE pairs after the input:

pounce problem.nl tol=1e-10 max_iter=500 print_level=8

From an options file — upstream ipopt.opt format:

pounce problem.nl --options-file ipopt.opt

Command-line KEY=VALUE pairs override values loaded from the options file.

Commonly used options

OptionMeaning
tolOverall convergence tolerance on the KKT error.
max_iterMaximum number of outer iterations.
print_levelConsole verbosity, 0 (silent) – 12 (maximum debug).
linear_solverKKT linear-solver backend. ma57 requires the ma57 feature build.
mu_strategyBarrier-parameter update strategy (monotone / adaptive).

For the full upstream option catalogue, see the Ipopt options reference; POUNCE reuses those names.

For scaling-specific options (nlp_scaling_method, target-gradient overrides, linear_system_scaling), see the Scaling reference page. For nonlinear bound tightening (presolve_fbbt, fbbt_tol, fbbt_max_iter, fbbt_max_constraints), see the FBBT reference page.

Barrier-parameter (μ) strategy

The barrier parameter μ controls the inner subproblem’s relaxation of complementarity. The two strategies are monotone (default — geometric schedule) and adaptive (quality-function oracle picks each μ from the current iterate’s complementarity). See μ-strategy for when to switch.

OptionDefaultMeaning
mu_strategymonotonemonotone (Fiacco–McCormick schedule) or adaptive (oracle-driven).
mu_oraclequality-functionAdaptive oracle: quality-function / loqo / probing.
mu_init0.1Seed value for μ at the first iterate.
mu_min1e-11Floor on μ; the solver stops decreasing past this.
mu_max1e5Cap on μ (adaptive mode). When set explicitly it overrides the mu_max_fact initialization.
mu_max_fact1e3Initializes mu_max as mu_max_fact · curr_avrg_compl at the first iterate (adaptive mode).
mu_target0.0Stop target for μ in monotone mode.
mu_linear_decrease_factor0.2κ_μ in μ ← min(κ_μ · μ, μ^θ_μ).
mu_superlinear_decrease_power1.5θ_μ in the same formula.
barrier_tol_factor10.0Inner-subproblem tolerance scales as barrier_tol_factor · μ.
sigma_max1e2Upper clamp on σ chosen by the quality-function oracle.
sigma_min1e-6Lower clamp on σ (raising this to 1e-2 can break a stair-stepping stall on some problems).
adaptive_mu_globalizationobj-constr-filterAdaptive-mode globalization: kkt-error, obj-constr-filter, or never-monotone-mode.

Quality-function oracle (adaptive-μ details)

These are only consumed when mu_strategy=adaptive and mu_oracle=quality-function. Defaults mirror upstream IpQualityFunctionMuOracle::RegisterOptions.

OptionDefaultMeaning
quality_function_norm_type2-norm-squaredNorm used to aggregate KKT components inside q(σ): 1-norm, 2-norm, 2-norm-squared, max-norm.
quality_function_centralitynoneCentrality penalty term: none, log, reciprocal, cubed-reciprocal.
quality_function_balancing_termnoneBalancing penalty when complementarity ≪ infeasibilities: none or cubic.
quality_function_max_section_steps8Cap on golden-section iterations when picking σ.
quality_function_section_sigma_tol1e-2Width tolerance in σ-space terminating the golden-section search.
quality_function_section_qf_tol0.0Relative flatness tolerance on q(σ) terminating golden section.

Adaptive-μ globalization

Tuning the safeguards that fall back to monotone-μ mode when the adaptive oracle stops making progress. Defaults mirror upstream IpAdaptiveMuUpdate::RegisterOptions.

OptionDefaultMeaning
adaptive_mu_safeguard_factor0.0LOQO safeguard floor on the oracle’s μ candidate.
adaptive_mu_monotone_init_factor0.8Multiplier on avrg_compl when seeding monotone mode after a bailout.
adaptive_mu_restore_previous_iteratenoRestore the latest free-mode iterate when switching to fixed mode.
adaptive_mu_kkterror_red_iters4Window length for the kkt-error globalization history.
adaptive_mu_kkterror_red_fact0.9999Required relative KKT-error reduction over that window.
adaptive_mu_kkt_norm_type2-norm-squaredNorm used to score the iterate in adaptive globalization decisions.

ℓ₁ penalty-barrier wrapper options

These tune the degenerate-NLP wrapper described in Running Solves. All are default-tuned and rarely need overriding:

OptionDefaultMeaning
l1_exact_penalty_barriernoRun the ℓ₁-exact penalty-barrier wrapper unconditionally.
l1_fallback_on_restoration_failurenoRetry with the wrapper only when the standard solve fails.
l1_penalty_init1.0Initial penalty weight ρ.
l1_penalty_max1e6Maximum penalty weight before declaring infeasibility.
l1_penalty_increase_factor8.0Multiplier applied to ρ each outer iteration.
l1_penalty_max_outer_iter8Maximum penalty outer iterations.
l1_slack_tol1e-6Slack tolerance for “constraints satisfied”.
l1_steering_factor10.0Steering-rule factor for ρ escalation.

NLP Presolve

POUNCE’s TNLP-wrapper presolve pipeline runs before the IPM starts. It tightens variable bounds, drops redundant rows, and (optionally) eliminates square auxiliary-equality sub-systems structurally. All are off by default — set the master switch first:

OptionDefaultMeaning
presolvenoMaster switch for the whole presolve layer. Off → wrapper is a no-op.
presolve_bound_tighteningyesPhase 1 — Andersen-style bound propagation from linear rows.
presolve_redundant_constraint_removalyesPhase 2 — drop linear constraints already implied by current bounds.
presolve_linear_eq_reductionnoPhase ≥2 — eliminate fixed singleton variables exposed by linear equalities.
presolve_licq_checkyesPhase 3 — detect rank-deficient equality blocks before the IPM starts.
presolve_licq_actionwarnWhat to do on degeneracy: warn (just report) or auto_l1 (turn on ℓ₁).
presolve_warm_z_boundsyesPhase 4 — warm-start bound multipliers when bounds get tightened by Phase 1.
presolve_bound_mult_init_val1.0Value used by Phase 4 for those warm-start hints.
presolve_max_passes3Fixed-point iteration cap across the bound-tightening passes.
presolve_print_level0Per-pass verbosity (0 silent, 5 per-pass, 8 per-transformation).

Feasibility-based bound tightening (Phase 1b)

Interval-arithmetic propagation through nonlinear constraint expression DAGs (see FBBT). Available today for .nl-loaded problems via NlTnlp; other TNLP sources opt out silently.

OptionDefaultMeaning
presolve_fbbtnoMaster switch. Requires presolve=yes and an ExpressionProvider.
fbbt_tol1e-6Minimum per-variable bound improvement to keep iterating.
fbbt_max_iter10Outer-sweep cap.
fbbt_max_constraints0Per-sweep cap on constraints inspected (0 = unlimited).

Auxiliary-equality preprocessing (Phase 0)

A separate set of options controls the structural elimination pass documented in Auxiliary-Equality Preprocessing:

OptionDefaultMeaning
presolve_auxiliarynoMaster switch for the Phase-0 structural elimination pass.
presolve_auxiliary_couplingsafeWhich coupling classes are eligible: none / safe / aggressive.
presolve_auxiliary_tol1e-8Residual tolerance for accepting a candidate block solve.
presolve_auxiliary_max_block_dim8Largest block the lightweight Newton solver will attempt (larger blocks rejected in v1).
presolve_auxiliary_wall_time_fraction0.1Fraction of the solver’s wall-time budget the pass is allowed to spend.
presolve_auxiliary_diagnosticsnoEmit the diagnostics summary via the journalist after Phase 0 runs.

FERAL backend tuning

linear_solver=feral (the default — see Commonly used options) is configurable through six feral_* options. Defaults are tuned for the IPM workload and rarely need changing; reach for these when profiling a specific problem.

OptionDefaultMeaning
feral_orderingautoFill-reducing ordering method (see table below). auto lets feral’s adaptive dispatcher pick per-matrix; auto_race measures the actual symbolic outcome and keeps the best.
feral_pivtol1e-8Relative Bunch-Kaufman partial-pivoting threshold u. Analog of ma27_pivtol / ma57_pivtol. Smaller → sparser L, faster, less stable; larger → more 2×2 blocks, denser, more stable. LAPACK’s textbook maximum-stability value is 0.5.
feral_refineyesIterative refinement on every back-solve. Closes the residual floor from cascade-break’s L-factor perturbation; disable only when timing the bare factor + back-solve in isolation.
feral_cascade_break(unset)Tri-state. Unset → inherit feral’s Phase B default (CB on with bounded delayed-pivot catchment). yes records explicit intent (no behavioural change). no reproduces pre-Phase-B behaviour by surfacing DelayBudgetExceeded on non-root cascade victims.
feral_fmanoDispatch dense kernels through fused multiply-add intrinsics. Roughly 2× throughput on aarch64 / x86_v3, at the cost of per-pivot rounding drift that trips more WrongInertia checks. Turn on when kernel throughput dominates and the IPM tolerates a noisier inertia signal.
feral_singular_pivot_floor1e-20Pounce’s analog of MA57’s CNTL(2). After a successful factor, the smallest accepted D-block pivot magnitude (scaled space) is compared against this absolute floor; if it falls below, the factor is reported Singular so the IPM bumps δ_w. 0 disables.

feral_ordering variants

All six concrete and adaptive options live under the same string option. feral_ordering also falls back to the POUNCE_FERAL_ORDERING environment variable when not set on the OptionsList.

ValueStrategy
autoDefault. Adaptive dispatcher: picks a concrete method per matrix from cheap pattern features. Branches: very-large-and-sparse (n > 100 000, avg degree < 5) → AMD; n ≤ 10 000 → AMF; otherwise → MetisND. One symbolic pass; right when the heuristic shape rules apply (the common case).
auto_raceRace-based dispatcher: runs full symbolic factorization on AMD, MetisND, ScotchND, KahipND and keeps the smallest factor_nnz. ~4× a single symbolic pass, paid once per problem (symbolic factorization is cached across numeric refactorizations with the same pattern). Use when the cheap dispatcher’s guess is suspect — e.g. pinene_3200_0009, where auto picks MetisND (88 s numeric factor) but amd factors in 19.5 s on the same matrix.
amdApproximate Minimum Degree (Amestoy/Davis/Duff). Pins AMD regardless of problem shape; robust default for IPM workloads. Best for very-large-and-sparse cases that the adaptive dispatcher already routes here.
amfApproximate Minimum Fill (HAMF4 variant of Amestoy 1999). Strong on small-and-sparse populations (n ≤ 10 000); aggregate fill ≈ 0.87× AMD on feral’s IPM small-sparse inventory.
metisferal-metis multilevel nested dissection. Tends to produce squarer fronts than AMD on banded / nearly-1D structure; preferred for large structured matrices.
scotchferal-scotch nested dissection. Similar regime to METIS; alternative when METIS is unavailable or for cross-validation.
kahipferal-kahip flow-based nested dissection with K1 preprocessing. Ties METIS on fill geomean at 4–6× per-call symbolic cost. Reach for it only when ND fill matters and per-call cost is amortized.

When in doubt: leave feral_ordering at the default. When a hard problem looks linear-solver-bound, try feral_ordering auto_race before per-variant manual sweeping — it’s the safe choice when the per-problem winner is uncertain.

Logging and colored output

POUNCE emits structured logs and a colored iteration table through the tracing ecosystem. Behavior is governed by environment variables (not solver options), so they apply to the pounce CLI, the C/Python frontends, and anything embedding the library.

VariableValuesEffect
RUST_LOGe.g. info, debug, pounce::restoration=debugLog verbosity / per-target filtering. Default info. Logs go to stderr.
POUNCE_LOG_FORMATtext (default) · jsonjson emits line-delimited JSON on stderr (incl. the per-iteration pounce::iteration stream) for Studio / CI ingestion.
NO_COLORset to any valueDisables ANSI color in the iteration table and logs (see https://no-color.org).
CLICOLOR_FORCEset to any valueForces color even when stdout is not a terminal.

Filtering by subsystem. Solver internals log under namespaced targets — pounce::algorithm, pounce::linsol, pounce::mu, pounce::sqp, pounce::linesearch, pounce::restoration, pounce::presolve, pounce::py. For example, to trace only the restoration phase:

RUST_LOG=pounce::restoration=debug pounce problem.nl

Program output vs. logs. The iteration table, the final summary, and --dump diagnostics are program output on stdout; diagnostic and progress messages are logs on stderr. Redirecting one does not affect the other:

pounce problem.nl > result.txt 2> solve.log

Color. The iteration table is colored with a tiger/rust theme: restoration lines take a background that varies by restoration kind (soft-stay → tan, soft-exit → amber, hard → deep rust), and the row text shades from black toward red as the primal step length alpha shrinks (stalling). Color is emitted only when stdout is a terminal; redirected output and NO_COLOR get plain text with identical column alignment.

Machine-readable iterations. POUNCE_LOG_FORMAT=json turns the per-iteration records into JSON on stderr:

POUNCE_LOG_FORMAT=json pounce problem.nl 2> iters.jsonl