# ripopt
[](https://crates.io/crates/ripopt)
[](https://pypi.org/project/ripopt/)
[](https://pypi.org/project/pyomo-ripopt/)
[](https://github.com/jkitchin/ripopt/actions/workflows/test.yml)
[](https://doi.org/10.5281/zenodo.19542664)
A memory-safe interior point optimizer written in Rust, inspired by [Ipopt](https://github.com/coin-or/Ipopt).
## What is this?
ripopt solves nonlinear programming (NLP) problems of the form:
```
min f(x)
s.t. g_l <= g(x) <= g_u
x_l <= x <= x_u
```
It implements a primal-dual interior point method with a barrier formulation, similar to the algorithm described in the Ipopt papers. The solver is written entirely in Rust (~21,700 lines) with no external C/Fortran dependencies.
## Features
- Primal-dual interior point method with logarithmic barrier
- Dense LDL^T factorization via Bunch-Kaufman pivoting with inertia detection
- Sparse multifrontal LDL^T factorization (via `rmumps` with SuiteSparse AMD ordering) for larger problems (n+m >= 110)
- Banded LDL^T solver for problems with detected-banded structure (e.g., PDE discretizations)
- Dense condensed KKT (Schur complement) for tall-narrow problems (m >> n, n <= 100)
- Sparse condensed KKT for reducing system size when m > 0
- Filter line search with switching condition and Armijo criterion
- Second-order corrections (SOC) for improved step acceptance
- **Mehrotra predictor-corrector** with Gondzio centrality corrections (enabled by default)
- Adaptive and monotone barrier parameter strategies with Mehrotra sigma-guided mu updates
- Fraction-to-boundary rule for primal and dual step sizes
- Support for equality constraints, inequality constraints, and variable bounds
- Warm-start initialization
- Two-phase restoration: fast Gauss-Newton + full NLP restoration subproblem
- Multi-attempt recovery with systematic barrier landscape perturbation
- Watchdog strategy for escaping narrow feasible corridors
- Automatic NE-to-LS reformulation for overdetermined nonlinear equation systems
- **Convergence polishing**: Newton polish for NE-to-LS, complementarity snap for IPM
- NLP scaling (gradient-based objective and constraint scaling)
- Local infeasibility detection for inconsistent constraint systems
- **Early stall detection**: bail out fast when stuck in early iterations to trigger fallbacks
- **Preprocessing**: Automatic elimination of fixed variables, redundant constraints, and bound tightening from single-variable linear constraints
- **Near-linear constraint detection**: Automatically identifies linear constraints and skips their Hessian contribution
- **Limited-memory Hessian approximation**: L-BFGS-in-IPM mode (`hessian_approximation_lbfgs`) replaces exact Hessian with L-BFGS curvature pairs, eliminating the need for second-derivative callbacks
- **Multi-solver fallback architecture**: L-BFGS, Augmented Lagrangian, SQP, and explicit slack reformulation
- **Parametric sensitivity analysis**: sIPOPT-style post-optimal sensitivity (`ds/dp = -M⁻¹ · Nₚ`) for computing how the optimal solution changes under parameter perturbations, plus reduced Hessian extraction for covariance estimation
- **C API** mirroring the Ipopt C interface for direct linking from C/C++/Python/Julia
- **AMPL NL interface** with Pyomo integration via `SolverFactory('ripopt')`, with `--help` listing all options. Supports AMPL **external functions** via the `funcadd_ASL` ABI (e.g. IDAES `cbrt`, custom property-package libraries) at solve time.
- **GAMS solver link** enabling `option nlp = ripopt;` in GAMS models via the GMO API
- **Julia/JuMP interface** (`Ripopt.jl`) via MathOptInterface, enabling `Model(Ripopt.Optimizer)` with full JuMP support
## Benchmarks
### Hock-Schittkowski Test Suite (120 problems)
| Metric | ripopt | Ipopt (native, MUMPS) |
|-----------------|--------------------|-----------------------|
| Problems solved | **118/120 (98.3%)**| 116/120 (96.7%) |
| Optimal | 118 | 116 |
| ripopt only | 2 | -- |
| Ipopt only | -- | 0 |
On 116 commonly-solved problems: **20.8x geometric mean speedup**, median 20.9x, ripopt faster on 114/116 (98%).
### CUTEst Benchmark Suite (727 problems)
| Metric | ripopt | Ipopt (C++ with MUMPS) |
|---------------|---------------------|------------------------|
| Total solved | 556/727 (76.5%) | 556/727 (76.5%) |
| Both solve | 521 | 521 |
| ripopt only | 38 | -- |
| Ipopt only | -- | 40 |
ripopt and native Ipopt are tied on CUTEst strict-Optimal at v0.7.1; v0.7.0 had ripopt at +4. The v0.7.0 → v0.7.1 cycle had churn of 20 regressions and 16 gains (net -4 problems), concentrated on least-squares / rank-deficient Jacobians; see CHANGELOG for details.
On 521 commonly-solved problems:
| Metric | Value |
|---------------------------------|------------------|
| Geometric mean speedup | **12.0x** |
| Median speedup | **21.0x** |
| Problems where ripopt is faster | 454/521 (87%) |
| ripopt 10x+ faster | 340/521 (65%) |
| Problems where Ipopt is faster | 67/521 (13%) |
**Interpreting the speed numbers.** Most CUTEst problems are small (n < 10) and solve in microseconds for ripopt, while Ipopt has a ~1-3ms floor from internal initialization. The per-iteration speedup on small problems comes from stack allocation, the absence of C/Fortran interop, and cache-efficient dense linear algebra. On larger problems, ripopt switches to sparse multifrontal LDL^T with SuiteSparse AMD ordering, and Ipopt's Fortran MUMPS has a per-factorization advantage. Ipopt uses fewer iterations on average on CUTEst (ripopt mean 62.3 vs Ipopt 39.5), reflecting its more mature barrier parameter tuning.
The speed advantage comes from:
1. **Lower per-iteration overhead.** ripopt's dense Bunch-Kaufman factorization avoids sparse symbolic analysis and has minimal allocation. For small-to-medium problems (n < 50), this gives 2-5x per-iteration speedup.
2. **Dense condensed KKT for tall-narrow problems.** When m >> n with n <= 100, ripopt reduces an (n+m)x(n+m) sparse factorization to an nxn dense solve, giving 100-800x speedup on problems like EXPFITC (n=5, m=502) and OET3 (n=4, m=1002).
3. **Mehrotra predictor-corrector with Gondzio corrections.** Enabled by default, reducing iteration counts on many problems.
4. **Fewer iterations on some problems.** NE-to-LS reformulation, two-phase restoration, and multi-solver fallback recover problems that Ipopt cannot solve.
Where Ipopt is faster:
1. **Large sparse problems.** Ipopt's Fortran MUMPS is ~10-15x faster per factorization than rmumps on 50K-100K systems.
2. **Some medium constrained problems.** A handful of problems (CORE1, HAIFAM, NET1) have high per-iteration cost in ripopt's line search or fallback cascade.
3. **Some difficult nonlinear problems.** Ipopt's mature barrier parameter tuning gives it an edge on specific hard problems.
### Large-Scale Benchmarks
Both solvers receive the exact same NlpProblem struct via the Rust trait interface, ensuring a fair comparison. ripopt uses rmumps (pure Rust multifrontal LDL^T with SuiteSparse AMD ordering); Ipopt uses MUMPS (Fortran).
| Problem | n | m | ripopt | time | Ipopt | time | speedup |
|-----------------|--------|--------|---------|---------|----------|---------|-----------|
| Rosenbrock 500 | 500 | 0 | Optimal | 0.003s | Optimal | 0.199s | **76.2x** |
| Bratu 1K | 1,000 | 998 | Optimal | 0.002s | Optimal | 0.002s | 1.1x |
| SparseQP 1K | 500 | 500 | Optimal | 0.176s | Optimal | 0.004s | 0.02x |
| OptControl 2.5K | 2,499 | 1,250 | Optimal | 0.006s | Optimal | 0.002s | 0.4x |
Numbers above are fresh (v0.7.0). Problems above 2.5K variables
(Poisson 2.5K, Rosenbrock 5K, Bratu 10K, OptControl 20K, Poisson 50K,
SparseQP 100K) are not re-run for v0.7.0: the stricter KKT
backward-error probe introduced in this release causes Poisson 2.5K
to exhaust `max_iter` instead of converging quickly, and the full
large-scale sweep is temporarily gated behind a separate investigation.
Historical numbers from v0.6.2 are preserved in
`benchmarks/large_scale/large_scale_results.txt` snapshots. ripopt's
advantage is strongest on unconstrained problems (L-BFGS fallback);
on large constrained problems Ipopt's Fortran MUMPS is ~10-15x faster
per factorization.
Run the benchmarks yourself: `make benchmark`
### Domain-Specific Benchmarks
| Suite | Problems | ripopt | Ipopt | Notes |
|------------------------------|----------|------------------|---------------|--------------------------------------------------------------------|
| Electrolyte thermodynamics | 13 | **13/13 (100%)** | 12/13 (92.3%) | 17.5x geo mean speedup; ripopt uniquely solves seawater speciation |
| Grid (AC Optimal Power Flow) | 4 | 3/4 (75%) | **4/4 (100%)**| 2.8x geo mean on 3 commonly-solved; case30_ieee regression |
| CHO parameter estimation | 1 | 0/1 | 0/1 | Large-scale (n=21,672, m=21,660); both hit iteration limit |
| Gas pipeline NLPs | 4 | see suite README | see suite README | PDE-discretized Euler equations on pipe networks (gaslib11/40, steady/dynamic). Standalone — does not feed `BENCHMARK_REPORT.md` |
| Water distribution NLPs | 6 | see suite README | see suite README | MINLPLib water network design instances (Hazen-Williams head-loss). Standalone — does not feed `BENCHMARK_REPORT.md` |
Run all benchmarks: `make benchmark`
## Installation
### Prerequisites: Rust and Cargo
ripopt is written in Rust. You need the Rust toolchain (compiler + Cargo build tool) installed.
**Install Rust via rustup** (the official installer, works on macOS, Linux, and WSL):
```bash
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
```
Follow the prompts (the defaults are fine). Then restart your shell or run:
```bash
source "$HOME/.cargo/env"
```
Verify the installation:
```bash
rustc --version
cargo --version
```
For other installation methods (Homebrew, distro packages, Windows), see the [official Rust installation guide](https://www.rust-lang.org/tools/install).
### Install ripopt
**From crates.io:**
```bash
cargo install ripopt
```
This installs the `ripopt` AMPL solver binary to `~/.cargo/bin/`.
To use ripopt as a library dependency in your Rust project:
```bash
cargo add ripopt
```
**From source** (clone the repository and run `make install`):
```bash
git clone https://github.com/jkitchin/ripopt.git
cd ripopt
make install
```
This does three things:
1. Builds the optimized release binary and shared library
2. Installs the `ripopt` AMPL solver binary to `~/.cargo/bin/` (which rustup already added to your `$PATH`)
3. Copies the shared library (`libripopt.dylib` on macOS, `libripopt.so` on Linux) to `~/.local/lib/`
> **PATH check:** The `ripopt` binary is installed to `~/.cargo/bin/`. If `ripopt --version` doesn't work after installation, make sure `~/.cargo/bin` is on your `$PATH` by adding this to your shell profile (`~/.bashrc`, `~/.zshrc`, etc.):
> ```bash
> export PATH="$HOME/.cargo/bin:$PATH"
> ```
> Then restart your shell or run `source ~/.bashrc` (or `~/.zshrc`).
> **Shared library:** If you use the C API, ensure `~/.local/lib` is in your library path:
> ```bash
> export LD_LIBRARY_PATH="$HOME/.local/lib:$LD_LIBRARY_PATH"
> ```
After installation, verify it works:
```bash
ripopt --version
```
### Using ripopt with Pyomo
Once `ripopt` is on your `$PATH` (the `make install` step above handles this), install the Pyomo solver plugin:
```bash
pip install ./pyomo-ripopt
```
This registers `ripopt` as a named solver with Pyomo's `SolverFactory`:
```python
from pyomo.environ import *
model = ConcreteModel()
# ... define your model ...
solver = SolverFactory('ripopt')
result = solver.solve(model, tee=True)
```
> **Note:** If you skip the `pip install` step, you can still use ripopt via the generic AMPL interface with `SolverFactory('asl:ripopt')`, as long as the `ripopt` binary is on your `$PATH`.
### Using ripopt as a Rust library
Add to your project's `Cargo.toml`:
```toml
[dependencies]
ripopt = { git = "https://github.com/jkitchin/ripopt" }
```
Or for a local checkout:
```toml
[dependencies]
ripopt = { path = "/path/to/ripopt" }
```
### Using the shared library (C/Python ctypes/Julia FFI)
After `make install`, the shared library is at `~/.local/lib/libripopt.dylib` (macOS) or `~/.local/lib/libripopt.so` (Linux). The C header `ripopt.h` is in the repository root.
```bash
# Compile a C program against the installed library
cc my_program.c -I/path/to/ripopt -L~/.local/lib -lripopt -lm
```
Or link directly from the build directory without installing:
```bash
cargo build --release
cc my_program.c -I. -Ltarget/release -lripopt \
-Wl,-rpath,$(pwd)/target/release -o my_program -lm
```
### Using ripopt with GAMS
ripopt includes a GAMS solver link (`gams/`) that bridges between GAMS's GMO API and ripopt's C API. This allows GAMS models to use ripopt as an NLP solver via `option nlp = ripopt;`.
**Build and install** (requires a GAMS installation):
```bash
cargo build --release
make -C gams
sudo make -C gams install # copies libs to GAMS sysdir, registers in gmscmpun.txt
```
**Use in a GAMS model:**
```gams
option nlp = ripopt;
Solve mymodel using nlp minimizing obj;
```
**Solver options** are set via a `ripopt.opt` file (same key-value format as Ipopt):
```
tol 1e-8
max_iter 1000
print_level 5
```
GAMS iteration and resource limits (`option iterlim`, `option reslim`) are automatically forwarded. The solver link supports NLP, DNLP, and RMINLP model types. When the analytical Hessian is not available (e.g., DNLP models), it automatically falls back to L-BFGS approximation.
**Test:**
```bash
sudo make -C gams test # solves HS071 and checks the result
```
### Using ripopt with Julia/JuMP
`Ripopt.jl` provides a [MathOptInterface](https://github.com/jump-dev/MathOptInterface.jl) (MOI) wrapper so ripopt can be used as a drop-in JuMP optimizer.
**Prerequisites:** Julia ≥ 1.9, JuMP, and the ripopt shared library.
**Install once** (adds Ripopt.jl to your global Julia environment):
```bash
cargo build --release # build libripopt.dylib / libripopt.so
julia -e '
import Pkg
Pkg.develop(path="Ripopt.jl") # or Pkg.add(url="...") for a remote install
'
```
**Use in a script or notebook:**
```julia
ENV["RIPOPT_LIBRARY_PATH"] = "/path/to/ripopt/target/release"
using JuMP, Ripopt
model = Model(Ripopt.Optimizer)
set_silent(model)
@variable(model, 1 <= x[1:4] <= 5)
set_start_value.(x, [1.0, 5.0, 5.0, 1.0])
@NLobjective(model, Min, x[1]*x[4]*(x[1]+x[2]+x[3]) + x[3])
@NLconstraint(model, x[1]*x[2]*x[3]*x[4] >= 25)
@NLconstraint(model, x[1]^2 + x[2]^2 + x[3]^2 + x[4]^2 == 40)
optimize!(model)
println(termination_status(model)) # LOCALLY_SOLVED
println(objective_value(model)) # ≈ 17.014
println(value.(x)) # ≈ [1.0, 4.743, 3.821, 1.379]
```
Or run the provided examples from the repo root:
```bash
RIPOPT_LIBRARY_PATH=target/release julia --project=@v1.12 Ripopt.jl/examples/jump_hs071.jl
RIPOPT_LIBRARY_PATH=target/release julia --project=@v1.12 Ripopt.jl/examples/jump_rosenbrock.jl
RIPOPT_LIBRARY_PATH=target/release julia --project=@v1.12 Ripopt.jl/examples/c_wrapper_hs071.jl
```
Solver options use the same names as Ipopt:
```julia
set_optimizer_attribute(model, "tol", 1e-10)
set_optimizer_attribute(model, "max_iter", 500)
set_optimizer_attribute(model, "mu_strategy", "adaptive")
set_time_limit_sec(model, 60.0)
```
Switching between ripopt and Ipopt requires only changing the optimizer constructor; the rest of the model is identical:
```julia
# With ripopt
model = Model(Ripopt.Optimizer)
# With Ipopt (if installed)
using Ipopt
model = Model(Ipopt.Optimizer)
```
### Uninstall
```bash
make uninstall
```
## Usage
### Defining a Problem
Implement the `NlpProblem` trait:
```rust
use ripopt::NlpProblem;
struct Rosenbrock;
impl NlpProblem for Rosenbrock {
fn num_variables(&self) -> usize { 2 }
fn num_constraints(&self) -> usize { 0 }
fn bounds(&self, x_l: &mut [f64], x_u: &mut [f64]) {
// Unconstrained: use infinity bounds
for i in 0..2 {
x_l[i] = f64::NEG_INFINITY;
x_u[i] = f64::INFINITY;
}
}
fn constraint_bounds(&self, _g_l: &mut [f64], _g_u: &mut [f64]) {}
fn initial_point(&self, x0: &mut [f64]) { x0[0] = -1.0; x0[1] = 1.0; }
fn objective(&self, x: &[f64], _new_x: bool) -> f64 {
100.0 * (x[1] - x[0] * x[0]).powi(2) + (1.0 - x[0]).powi(2)
}
fn gradient(&self, x: &[f64], _new_x: bool, grad: &mut [f64]) {
grad[0] = -400.0 * x[0] * (x[1] - x[0] * x[0]) - 2.0 * (1.0 - x[0]);
grad[1] = 200.0 * (x[1] - x[0] * x[0]);
}
fn constraints(&self, _x: &[f64], _g: &mut [f64]) {}
fn jacobian_structure(&self) -> (Vec<usize>, Vec<usize>) { (vec![], vec![]) }
fn jacobian_values(&self, _x: &[f64], _vals: &mut [f64]) {}
fn hessian_structure(&self) -> (Vec<usize>, Vec<usize>) {
(vec![0, 1, 1], vec![0, 0, 1]) // lower triangle
}
fn hessian_values(&self, x: &[f64], obj_factor: f64, _lambda: &[f64], vals: &mut [f64]) {
vals[0] = obj_factor * (-400.0 * (x[1] - 3.0 * x[0] * x[0]) + 2.0);
vals[1] = obj_factor * (-400.0 * x[0]);
vals[2] = obj_factor * 200.0;
}
}
```
### Solving
```rust
use ripopt::{SolverOptions, solve};
let problem = Rosenbrock;
let options = SolverOptions::default();
let result = solve(&problem, &options);
println!("Status: {:?}", result.status);
println!("Objective: {:.6e}", result.objective);
println!("Solution: {:?}", result.x);
println!("Iterations: {}", result.iterations);
```
### Solver Options
Key options (all have Ipopt-matching defaults):
| Option | Default | Description |
|---------------------------------|---------|------------------------------------------------------------|
| `tol` | 1e-8 | Convergence tolerance |
| `max_iter` | 3000 | Maximum iterations |
| `mu_init` | 0.1 | Initial barrier parameter |
| `print_level` | 5 | Output verbosity (0=silent, 5=verbose) |
| `mu_strategy_adaptive` | true | Adaptive vs monotone barrier update |
| `max_soc` | 4 | Maximum second-order correction steps |
| `max_wall_time` | 0.0 | Wall-clock time limit in seconds (0=no limit) |
| `warm_start` | false | Enable warm-start initialization |
| `constr_viol_tol` | 1e-4 | Constraint violation tolerance |
| `dual_inf_tol` | 1.0 | Dual infeasibility tolerance |
| `enable_preprocessing` | true | Eliminate fixed variables and redundant constraints |
| `detect_linear_constraints` | true | Skip Hessian for linear constraints |
| `enable_sqp_fallback` | true | SQP fallback for constrained problems |
| `hessian_approximation_lbfgs` | false | Use L-BFGS Hessian approximation (no exact Hessian needed) |
| `enable_lbfgs_hessian_fallback` | true | Auto-retry with L-BFGS Hessian when exact Hessian fails |
| `mehrotra_pc` | true | Mehrotra predictor-corrector for better centering |
| `gondzio_mcc_max` | 3 | Maximum Gondzio centrality corrections per iteration |
| `early_stall_timeout` | 120.0 | Max seconds for first 3 iterations (0=off) |
| `linear_solver` | direct | KKT solver: direct, iterative (MINRES), or hybrid |
### Result
`SolveResult` contains:
- `x` -- optimal primal variables
- `objective` -- optimal objective value f(x*)
- `constraint_multipliers` -- Lagrange multipliers for constraints (y)
- `bound_multipliers_lower` / `bound_multipliers_upper` -- bound multipliers (z_L, z_U)
- `constraint_values` -- constraint values g(x*)
- `status` -- one of: `Optimal`, `Infeasible`, `LocalInfeasibility`, `MaxIterations`, `NumericalError`, `Unbounded`, `RestorationFailed`, `InternalError`
- `iterations` -- number of IPM iterations
## C API
ripopt exposes a C API that mirrors the [Ipopt C interface](https://coin-or.github.io/Ipopt/INTERFACES.html#INTERFACE_C), enabling direct linking from C, C++, Python (`ctypes`/`cffi`), Julia, and any language with C FFI support — without the subprocess/file overhead of the NL interface. If you have existing Ipopt C code, migrating to ripopt requires only header/function renaming; the callback signatures are identical.
### Build the shared library
```bash
cargo build --release
# produces target/release/libripopt.dylib (macOS) or libripopt.so (Linux)
```
### C header
Include `ripopt.h` (repo root) in your C project. It defines version macros, callback typedefs, return status codes, and all public functions:
```c
#include "ripopt.h"
// Check version at compile time
printf("ripopt %s\n", RIPOPT_VERSION); // "0.7.1"
```
### Callback signatures
The five callback types are identical to the Ipopt C interface. All callbacks return `1` on success, `0` on error (the solver will abort if a callback returns `0`):
```c
typedef int (*Eval_F_CB) (int n, const double *x, int new_x,
double *obj_value, void *user_data);
typedef int (*Eval_Grad_F_CB)(int n, const double *x, int new_x,
double *grad_f, void *user_data);
typedef int (*Eval_G_CB) (int n, const double *x, int new_x,
int m, double *g, void *user_data);
typedef int (*Eval_Jac_G_CB)(int n, const double *x, int new_x,
int m, int nele_jac,
int *iRow, int *jCol, double *values,
void *user_data);
typedef int (*Eval_H_CB) (int n, const double *x, int new_x,
double obj_factor,
int m, const double *lambda, int new_lambda,
int nele_hess,
int *iRow, int *jCol, double *values,
void *user_data);
```
**Two-call protocol for Jacobian and Hessian:** When `values == NULL`, fill `iRow`/`jCol` with the sparsity pattern (0-based indexing); when `values != NULL`, fill numerical values in the same element order as the pattern. The Hessian uses the **lower triangle** only.
**Sign convention:** ripopt uses the Ipopt convention L = f(x) + y^T g(x). The Hessian callback receives `obj_factor` and `lambda` and should compute `obj_factor * ∇²f + Σ lambda[i] * ∇²g_i`.
### Lifecycle
```c
#include "ripopt.h"
// 1. Create handle
RipoptProblem nlp = ripopt_create(
n, x_l, x_u, // variable bounds (use ±1e30 for ±∞)
m, g_l, g_u, // constraint bounds (g_l == g_u for equality)
nele_jac, nele_hess, // number of nonzeros
eval_f, eval_grad_f, eval_g, eval_jac_g, eval_h);
// 2. Set options (Ipopt-compatible key names)
ripopt_add_int_option(nlp, "print_level", 5);
ripopt_add_num_option(nlp, "tol", 1e-8);
ripopt_add_str_option(nlp, "mu_strategy", "adaptive");
// 3. Solve (x: in = initial point, out = solution)
double obj_val;
int status = ripopt_solve(nlp, x, NULL, &obj_val,
NULL, NULL, NULL, NULL);
// status == 0 → RIPOPT_SOLVE_SUCCEEDED
// 4. Free
ripopt_free(nlp);
```
For unconstrained problems, pass `m=0` and `NULL` for `g_l`/`g_u`.
**Infinity bounds:** Use `HUGE_VAL` (from `<math.h>`) for "no bound". Internally, any value beyond `±1e19` is treated as unbounded. Avoid using finite large values like `1e30` — they may cause numerical issues.
### Extracting multipliers
All output pointers except `x` are optional (pass `NULL` to skip). Here is how to extract the full solution including Lagrange multipliers and bound multipliers:
```c
double x[4] = {1.0, 5.0, 5.0, 1.0}; // initial point
double obj_val = 0.0;
double g[2] = {0.0, 0.0}; // constraint values at solution
double mult_g[2] = {0.0, 0.0}; // constraint multipliers (lambda)
double mult_xl[4]= {0.0, 0.0, 0.0, 0.0}; // lower bound multipliers (z_L)
double mult_xu[4]= {0.0, 0.0, 0.0, 0.0}; // upper bound multipliers (z_U)
int status = ripopt_solve(nlp, x, g, &obj_val,
mult_g, mult_xl, mult_xu,
NULL); // user_data
// At the solution:
// - x[] contains the optimal primal variables
// - obj_val is f(x*)
// - g[] contains g(x*) — verify constraints are satisfied
// - mult_g[] contains the Lagrange multipliers for constraints
// (nonzero for active constraints)
// - mult_xl[] contains z_L (positive when x is at its lower bound)
// - mult_xu[] contains z_U (positive when x is at its upper bound)
```
The `user_data` pointer is forwarded to every callback unchanged — use it to pass problem-specific data (e.g., model parameters) without globals.
### Return status
| Code | Enum constant | Meaning |
|------|----------------------------------------|-------------------------------------------------|
| 0 | `RIPOPT_SOLVE_SUCCEEDED` | Converged to optimal solution |
| 2 | `RIPOPT_INFEASIBLE_PROBLEM` | Problem is locally infeasible |
| 5 | `RIPOPT_MAXITER_EXCEEDED` | Reached iteration limit |
| 6 | `RIPOPT_RESTORATION_FAILED` | Feasibility restoration failed |
| 7 | `RIPOPT_ERROR_IN_STEP_COMPUTATION` | Numerical difficulties |
| 10 | `RIPOPT_NOT_ENOUGH_DEGREES_OF_FREEDOM` | Problem has too few free variables |
| 11 | `RIPOPT_INVALID_PROBLEM_DEFINITION` | Problem appears unbounded |
| -1 | `RIPOPT_INTERNAL_ERROR` | Internal error |
Status 0 indicates a successful solve. All others indicate failure — check your problem formulation, initial point, or try adjusting options.
### Options reference
Option-setting functions return `1` on success, `0` if the keyword is unknown. All option keywords match Ipopt naming conventions.
**Numeric options** (`ripopt_add_num_option`):
| Option | Default | Description |
|------------------------------|---------|-------------------------------------------------|
| `tol` | 1e-8 | Convergence tolerance |
| `mu_init` | 0.1 | Initial barrier parameter |
| `mu_min` | 1e-11 | Minimum barrier parameter |
| `bound_push` | 1e-2 | Initial bound push |
| `bound_frac` | 1e-2 | Initial bound fraction |
| `constr_viol_tol` | 1e-4 | Constraint violation tolerance |
| `dual_inf_tol` | 1.0 | Dual infeasibility tolerance |
| `compl_inf_tol` | 1e-4 | Complementarity tolerance |
| `max_wall_time` | 0.0 | Wall-clock time limit in seconds (0 = no limit) |
| `warm_start_bound_push` | 1e-3 | Warm-start bound push |
| `warm_start_bound_frac` | 1e-3 | Warm-start bound fraction |
| `warm_start_mult_bound_push` | 1e-3 | Warm-start multiplier push |
| `nlp_lower_bound_inf` | -1e20 | Threshold for -infinity bounds |
| `nlp_upper_bound_inf` | 1e20 | Threshold for +infinity bounds |
| `kappa` | 10.0 | Adaptive mu divisor |
| `constr_mult_init_max` | 1000.0 | Max initial constraint multiplier |
| `barrier_tol_factor` | 10.0 | Barrier tolerance factor |
**Integer options** (`ripopt_add_int_option`):
| Option | Default | Description |
|------------------------|---------|--------------------------------------------------------|
| `max_iter` | 3000 | Maximum iterations |
| `print_level` | 5 | Output verbosity (0 = silent, 5 = verbose, 12 = debug) |
| `max_soc` | 4 | Maximum second-order correction steps |
| `sparse_threshold` | 110 | KKT dimension threshold for sparse solver |
| `restoration_max_iter` | 200 | Max iterations in NLP restoration subproblem |
**String options** (`ripopt_add_str_option`):
| Option | Default | Values | Description |
|---------------------------------|--------------|-------------------------------|---------------------------------------------------|
| `mu_strategy` | `"adaptive"` | `"adaptive"`, `"monotone"` | Barrier parameter update strategy |
| `warm_start_init_point` | `"no"` | `"yes"`, `"no"` | Enable warm-start initialization |
| `mu_allow_increase` | `"yes"` | `"yes"`, `"no"` | Allow barrier parameter increase |
| `least_squares_mult_init` | `"yes"` | `"yes"`, `"no"` | LS estimate for initial multipliers |
| `enable_slack_fallback` | `"yes"` | `"yes"`, `"no"` | Slack reformulation fallback |
| `enable_lbfgs_fallback` | `"yes"` | `"yes"`, `"no"` | L-BFGS fallback for unconstrained |
| `enable_al_fallback` | `"yes"` | `"yes"`, `"no"` | Augmented Lagrangian fallback |
| `enable_preprocessing` | `"yes"` | `"yes"`, `"no"` | Preprocessing (fixed vars, redundant constraints) |
| `detect_linear_constraints` | `"yes"` | `"yes"`, `"no"` | Skip Hessian for linear constraints |
| `enable_sqp_fallback` | `"yes"` | `"yes"`, `"no"` | SQP fallback for constrained problems |
| `hessian_approximation` | `"exact"` | `"exact"`, `"limited-memory"` | Use L-BFGS Hessian approximation |
| `enable_lbfgs_hessian_fallback` | `"yes"` | `"yes"`, `"no"` | Auto-retry with L-BFGS Hessian on failure |
### Error handling
- **Callback errors:** If any callback returns `0`, the solver aborts and returns `RIPOPT_ERROR_IN_STEP_COMPUTATION` (7). Always return `1` from callbacks unless you detect a problem (e.g., NaN in inputs).
- **Unknown options:** `ripopt_add_*_option` returns `0` for unrecognized keywords. Check the return value if you want to detect typos.
- **NULL safety:** `ripopt_free(NULL)` is a no-op (safe to call). All output pointers in `ripopt_solve` except `x` may be `NULL`.
- **Memory:** The problem handle owns all internal memory. Call `ripopt_free()` once when done. Do not use the handle after freeing.
### Migrating from Ipopt
If you have existing Ipopt C code, the migration is straightforward:
1. **Header:** `#include "IpStdCInterface.h"` → `#include "ripopt.h"`
2. **Handle type:** `IpoptProblem` → `RipoptProblem` (both are `void*`)
3. **Functions:** Rename `CreateIpoptProblem` → `ripopt_create`, `FreeIpoptProblem` → `ripopt_free`, `AddIpoptNumOption` → `ripopt_add_num_option`, etc.
4. **Callbacks:** No changes required — signatures are identical
5. **Status codes:** Similar semantics but different enum names (e.g., `Solve_Succeeded` → `RIPOPT_SOLVE_SUCCEEDED`)
6. **Infinity:** Ipopt uses ±2e19 by default; ripopt uses ±1e30 in bounds and ±1e20 for `nlp_*_bound_inf`
7. **Linking:** `-lipopt` → `-lripopt`
### Compile and run the examples
```bash
cargo build --release
# HS071 — constrained NLP with inequality + equality constraints
cc examples/c_api_test.c -I. -Ltarget/release -lripopt \
-Wl,-rpath,$(pwd)/target/release -o c_api_test -lm
./c_api_test
# Rosenbrock — unconstrained optimization
cc examples/c_rosenbrock.c -I. -Ltarget/release -lripopt \
-Wl,-rpath,$(pwd)/target/release -o c_rosenbrock -lm
./c_rosenbrock
# HS035 — bound-constrained QP with inequality
cc examples/c_hs035.c -I. -Ltarget/release -lripopt \
-Wl,-rpath,$(pwd)/target/release -o c_hs035 -lm
./c_hs035
# Full multiplier extraction and options demonstration
cc examples/c_example_with_options.c -I. -Ltarget/release -lripopt \
-Wl,-rpath,$(pwd)/target/release -o c_example_with_options -lm
./c_example_with_options
```
## Automatic Differentiation
Implementing `NlpProblem` requires hand-derived gradients, Jacobians, and Hessians, which is tedious and error-prone for real-world problems. The [ipopt-ad](https://github.com/prehner/ipopt-ad) crate (by [@prehner](https://github.com/prehner)) eliminates this by computing exact derivatives automatically via the [num-dual](https://crates.io/crates/num-dual) crate. Users write their objective and constraints once as generic Rust functions, and ipopt-ad handles all derivative computation and sparsity detection.
With automatic differentiation, the HS071 objective and constraints reduce to:
```rust
fn objective<D: DualNum<f64> + Copy>(x: SVector<D, 4>) -> D {
x[0] * x[3] * (x[0] + x[1] + x[2]) + x[2]
}
fn constraints<D: DualNum<f64> + Copy>(x: SVector<D, 4>) -> SVector<D, 2> {
SVector::from([
x[0] * x[1] * x[2] * x[3],
x[0] * x[0] + x[1] * x[1] + x[2] * x[2] + x[3] * x[3],
])
}
```
No gradient, Jacobian, or Hessian code needed. See `examples/autodiff_num_dual.rs` for a complete working example:
```bash
cargo run --example autodiff_num_dual --features num-dual
```
## Examples
### Rust
```bash
# Rosenbrock function (unconstrained with bounds)
cargo run --example rosenbrock
# HS071 (constrained NLP with inequalities)
cargo run --example hs071
# HS071 with automatic differentiation (no hand-derived derivatives)
cargo run --example autodiff_num_dual --features num-dual
# Benchmark timing across 5 problems
cargo run --release --example benchmark
# Parametric sensitivity analysis
cargo run --release --example sensitivity
```
### C
See [Compile and run the examples](#compile-and-run-the-examples) above for build instructions. The C examples are:
| Example | Problem type | Demonstrates |
|----------------------------|-----------------------------|--------------------------------------------------------------|
| `c_api_test.c` | HS071 (constrained) | Basic usage, all 5 callbacks |
| `c_rosenbrock.c` | Rosenbrock (unconstrained) | No constraints, no bounds |
| `c_hs035.c` | HS035 (bounds + inequality) | Bound multipliers, constraint multiplier |
| `c_example_with_options.c` | HS071 (multiple solves) | Options tuning, multiplier extraction, status interpretation |
## Tests
```bash
cargo test
```
230 tests total:
- **131 unit tests**: Dense LDL factorization, convergence checking, filter line search, fraction-to-boundary, KKT assembly, restoration, preprocessing, linearity detection, SQP, linear solver, autodiff, L-BFGS, sensitivity analysis
- **12 C API tests**: FFI integration tests
- **29 integration tests**: Rosenbrock, SimpleQP, HS071, HS035, PureBoundConstrained, MultipleEqualityConstraints, NE-to-LS reformulation, augmented Lagrangian, NL file parsing, IPM code paths, parametric sensitivity, and more
- **15 HS regression tests**: Selected Hock-Schittkowski problems for regression checking
- **14 coverage tests**: Augmented Lagrangian convergence paths, NL parser/solver pipeline, autodiff tape operations, IPM preprocessing/condensed KKT/unbounded detection
## Code Coverage
ripopt uses `cargo-llvm-cov` for code coverage measurement:
```bash
# Run tests with coverage and print summary
cargo llvm-cov test
# Detailed line-by-line report
cargo llvm-cov test --text
# HTML report (opens in browser)
cargo llvm-cov test --html && open target/llvm-cov/html/index.html
```
Current coverage by module:
| Module | Line Coverage |
|--------------------------|---------------|
| options.rs | 100% |
| slack_formulation.rs | 97% |
| warmstart.rs | 97% |
| filter.rs | 95% |
| restoration_nlp.rs | 93% |
| convergence.rs | 92% |
| sensitivity.rs | 91% |
| dense.rs (linear solver) | 91% |
| banded.rs | 91% |
| nl/header.rs | 91% |
| sqp.rs | 90% |
| preprocessing.rs | 85% |
| linear_solver/mod.rs | 84% |
| sparse.rs | 82% |
| kkt.rs | 81% |
| restoration.rs | 79% |
| multifrontal.rs | 78% |
| nl/autodiff.rs | 74% |
| lbfgs.rs | 70% |
| nl/parser.rs | 70% |
| augmented_lagrangian.rs | 66% |
| c_api.rs | 61% |
| nl/problem_impl.rs | 58% |
| linearity.rs | 53% |
| ipm.rs | 52% |
| nl/expr.rs | 37% |
| hybrid.rs | 0% |
| iterative.rs | 0% |
**Overall: 62% line coverage** (248 tests)
The `hybrid.rs` and `iterative.rs` modules implement opt-in linear
solvers (`LinearSolverChoice::Hybrid`, `::Iterative`) that no test
currently exercises — they are never selected by the default solver
path. Their coverage is reported as 0% until dedicated tests land.
## Architecture
```
src/
lib.rs Public API (solve function, re-exports)
c_api.rs C FFI layer (extern "C" functions, ripopt.h)
problem.rs NlpProblem trait definition
options.rs SolverOptions with Ipopt-matching defaults
result.rs SolveResult and SolveStatus
ipm.rs Main IPM loop, barrier updates, line search, NE-to-LS detection, NLP scaling
kkt.rs KKT system assembly, solution, and inertia correction
convergence.rs Convergence checking (primal/dual/complementarity)
filter.rs Filter line search mechanism
restoration.rs Gauss-Newton restoration phase with adaptive LM regularization
restoration_nlp.rs Full NLP restoration subproblem (Phase 2)
lbfgs.rs L-BFGS solver for unconstrained/bound-constrained problems
augmented_lagrangian.rs Augmented Lagrangian fallback for constrained problems
sqp.rs SQP fallback for constrained problems
sensitivity.rs Parametric sensitivity analysis (sIPOPT-style)
slack_formulation.rs Explicit slack reformulation fallback
preprocessing.rs Fixed variable elimination, redundant constraint removal, bound tightening
linearity.rs Near-linear constraint detection
warmstart.rs Warm-start initialization
linear_solver/
mod.rs LinearSolver trait, SymmetricMatrix, KktMatrix
dense.rs Dense LDL^T (Bunch-Kaufman) factorization
banded.rs Banded LDL^T for problems with small bandwidth
multifrontal.rs Multifrontal sparse LDL^T via rmumps (default, SuiteSparse AMD ordering)
sparse.rs Sparse LDL^T via faer (optional)
iterative.rs MINRES iterative solver with incomplete LDL^T preconditioner
hybrid.rs Hybrid direct/iterative solver with automatic switching
tests/
correctness.rs Integration tests (22 NLP problems)
ipm_paths.rs IPM code path tests (condensed KKT, unbounded, NE-to-LS, preprocessing)
sensitivity.rs Parametric sensitivity integration tests
hs_regression.rs HS suite regression tests (15 problems)
c_api.rs C API integration tests (12 tests via FFI)
lbfgs_ipm.rs L-BFGS Hessian approximation tests
iterative_solvers.rs Iterative/hybrid solver tests
large_scale.rs Large-scale correctness tests (up to 100K variables)
large_scale_benchmark.rs Large-scale ripopt vs Ipopt comparison
nl_integration.rs NL file parsing and solving tests
gams/
gams_ripopt.c GAMS solver link (GMO API → ripopt C API bridge)
Makefile Build, install, and test targets
install.sh Registration script for gmscmpun.txt
test_hs071.gms Smoke test (HS071 via `option nlp = ripopt`)
examples/
rosenbrock.rs Unconstrained optimization
hs071.rs Constrained NLP
sensitivity.rs Parametric sensitivity analysis demo
benchmark.rs Timing benchmark
c_api_test.c HS071 via the C API
c_rosenbrock.c Unconstrained Rosenbrock via C API
c_hs035.c Bound-constrained QP via C API
c_example_with_options.c Options and multiplier extraction demo
```
## Algorithm Details
### Preprocessing
Before solving, ripopt automatically analyzes the problem to reduce its size:
1. **Fixed variable elimination**: Variables with `x_l == x_u` are removed and set to their fixed values in all evaluations
2. **Redundant constraint removal**: Duplicate constraints (same Jacobian structure, values, and bounds) are eliminated
3. **Bound tightening**: Single-variable linear constraints are used to tighten variable bounds
The reduced problem is solved and the solution is mapped back to the original dimensions. Disable with `enable_preprocessing: false`.
### Near-Linear Constraint Detection
The Jacobian is evaluated at two points to identify linear constraints (where all Jacobian entries remain constant). For linear constraints, the Hessian contribution `lambda[i] * nabla^2 g_i` is exactly zero and is skipped, reducing computation in the Hessian evaluation.
### Core Interior Point Method
The solver follows the primal-dual barrier method from the Ipopt papers (Wachter & Biegler, 2006). At each iteration it:
1. Assembles and factors the KKT system using dense LDL^T (Bunch-Kaufman), sparse multifrontal LDL^T (rmumps), or dense condensed Schur complement for tall-narrow problems
2. Computes inertia of the factorization and applies regularization if needed
3. Applies Mehrotra predictor-corrector with Gondzio centrality corrections (default on)
4. Computes search directions with iterative refinement (up to 3 rounds)
5. Applies second-order corrections (SOC) when the initial step is rejected
6. Uses a filter line search with backtracking to ensure sufficient progress
7. Updates the barrier parameter adaptively using Mehrotra sigma-guided updates
### Multi-Solver Fallback Architecture
When the primary IPM fails, ripopt automatically tries alternative solvers:
1. **L-BFGS**: Tried first for unconstrained problems (m=0, no bounds); used as fallback for bound-constrained problems
2. **L-BFGS Hessian approximation**: Retries the IPM with L-BFGS curvature pairs replacing the exact Hessian (helps when the Hessian is ill-conditioned or buggy)
3. **Augmented Lagrangian**: PHR penalty method for constrained problems, with the IPM solving each AL subproblem
4. **SQP**: Equality-constrained Sequential Quadratic Programming with l1 merit function line search
5. **Explicit slack reformulation**: Converts g(x) to g(x)-s=0 with bounds on s, stabilizing multiplier oscillation at degenerate points
6. **Best-du tracking**: Throughout the solve, tracks the iterate with lowest dual infeasibility and recovers it at max iterations
### Limited-Memory Hessian Approximation (L-BFGS-in-IPM)
When `hessian_approximation_lbfgs = true`, the IPM replaces exact Hessian evaluations with an L-BFGS curvature approximation. This eliminates the need for `hessian_values()` callbacks entirely.
**How it works:**
1. Each IPM iteration, after accepting a step, the solver computes `s_k = x_{k+1} - x_k` and `y_k = ∇L(x_{k+1}) - ∇L(x_k)` from Lagrangian gradient differences
2. Powell damping ensures positive curvature (`s^T y > 0`)
3. An explicit dense B_k matrix is formed from the L-BFGS pairs and used in KKT assembly
4. Up to 10 curvature pairs are stored (O(n·m) memory where m=10)
**When to use it:**
- Neural networks in NLPs (dense Hessian, O(n²) memory prohibitive)
- Problems where second derivatives are unavailable or expensive
- Rapid prototyping (skip Hessian derivation)
- When the exact Hessian is ill-conditioned or buggy
**Automatic fallback:** By default (`enable_lbfgs_hessian_fallback = true`), if the exact-Hessian IPM fails with MaxIterations, NumericalError, or RestorationFailed, the solver automatically retries with L-BFGS Hessian approximation. This helps when the user-provided Hessian is inaccurate.
**Example:**
```rust
let options = SolverOptions {
hessian_approximation_lbfgs: true,
..SolverOptions::default()
};
let result = ripopt::solve(&problem, &options);
```
See `examples/lbfgs_hessian.rs` for complete working examples.
### Parametric Sensitivity Analysis
ripopt provides sIPOPT-style post-optimal sensitivity analysis: after solving an NLP, compute how the optimal solution changes when problem parameters are perturbed, without re-solving. This is useful for:
- **What-if analysis**: How does the optimal design change if a constraint bound shifts?
- **Uncertainty propagation**: Map parameter uncertainty to solution uncertainty via the reduced Hessian
- **Real-time optimization**: Update the solution for small disturbances at near-zero cost
The core equation is `ds/dp = -M⁻¹ · Nₚ` — one backsolve using the already-factored KKT matrix.
**Usage**: Implement the `ParametricNlpProblem` trait (extends `NlpProblem` with parameter derivative methods), then call `solve_with_sensitivity()`:
```rust
use ripopt::{ParametricNlpProblem, SolverOptions};
// Implement ParametricNlpProblem for your problem type...
// (adds num_parameters, jacobian_p_*, hessian_xp_* methods)
let mut ctx = ripopt::solve_with_sensitivity(&problem, &options);
// Compute sensitivity for a parameter perturbation Δp
let dp = [0.1]; // perturbation vector
let sens = ctx.compute_sensitivity(&problem, &[&dp]).unwrap();
// Predict perturbed solution: x(p + Δp) ≈ x* + dx
let x_new: Vec<f64> = ctx.result.x.iter()
.zip(sens.dx_dp[0].iter())
.map(|(x, dx)| x + dx)
.collect();
// Extract reduced Hessian for covariance estimation
let cov = ctx.reduced_hessian().unwrap();
```
On the HS071 problem with a parametric constraint bound, the sensitivity prediction matches re-solve to within 1e-5:
```
Predicted x(p=40.1): (1.000000, 4.751642, 3.824904, 1.375540)
Actual x(p=40.1): (1.000000, 4.751634, 3.824896, 1.375553)
Prediction errors: 8.4e-6, 8.0e-6, 1.4e-5
```
See `examples/sensitivity.rs` for a complete working example.
### Condensed KKT System
For problems where the number of constraints m exceeds 2n, the solver automatically uses a condensed (Schur complement) formulation. This reduces the factorization cost from O((n+m)^3) to O(n^2 m + n^3), enabling efficient handling of problems with many constraints and few variables.
### NE-to-LS Reformulation
When the solver detects an overdetermined nonlinear equation system (m >= n, f(x) = 0, all equality constraints, starting point not already feasible), it automatically reformulates the problem as unconstrained least-squares minimization:
```
min 0.5 * ||g(x) - target||^2
```
using a full Hessian (J^T J + sum of r_i * nabla^2 g_i). If the residual is small at the solution, the original system is consistent and `Optimal` is reported. Otherwise, `LocalInfeasibility` is reported with the best least-squares solution.
### Two-Phase Restoration
When the filter line search fails:
1. **Phase 1 (Gauss-Newton)**: Fast feasibility solver minimizing ||violation||^2 with gradient descent fallback
2. **Phase 2 (NLP restoration)**: Full barrier subproblem with p/n slack variables (Ipopt formulation)
3. **Multi-attempt recovery**: Up to 6 attempts cycling barrier parameter perturbations [10x, 0.1x, 100x, 0.01x, 1000x, 0.001x] with x perturbation
### Watchdog Strategy
The solver implements a watchdog mechanism that temporarily relaxes the filter acceptance criteria when progress stalls due to shortened steps. This helps escape narrow feasible corridors where strict Armijo conditions are too conservative.
## Profiling
### Built-in Phase Timing
ripopt includes per-iteration phase timing instrumentation. When `print_level >= 5` (the default), a summary table is printed at the end of each solve showing where CPU time is spent:
```
Phase breakdown (47 iterations):
Problem eval 0.234s (45.2%)
KKT assembly 0.089s (17.2%)
Factorization 0.156s (30.1%)
Direction solve 0.012s (2.3%)
Line search 0.021s (4.1%)
Other 0.006s (1.1%)
Total 0.518s
```
To suppress timing output, set `print_level: 0` in `SolverOptions`.
### External Profiling with samply
Release builds include debug symbols (`debug = true` in `[profile.release]`), so external profilers can show function names. [samply](https://github.com/mstange/samply) provides flamegraph visualization on macOS and Linux:
```bash
cargo install samply
cargo build --release --bin hs_suite
samply record target/release/hs_suite
```
This opens a Firefox Profiler UI in the browser with a full call tree and flamegraph. Look for wide bars under `solve_ipm` to identify dominant functions.
On macOS, Instruments (Xcode) also works without any additional setup:
```bash
cargo build --release --bin hs_suite
xcrun xctrace record --template "Time Profiler" --launch target/release/hs_suite
```
## Sign Convention
ripopt uses the Ipopt convention where the Lagrangian is:
```
L = f(x) + y^T g(x)
```
For inequality constraints `g(x) >= g_l`, the multiplier `y` is negative at optimality. For equality constraints, `y` can be positive or negative.
## License
[EPL-2.0](LICENSE) (Eclipse Public License 2.0), consistent with Ipopt.
The `rmumps` workspace member (pure Rust multifrontal solver) is licensed under [CeCILL-C](rmumps/LICENSE), a LGPL-compatible free software license.