pdlp

Primal-Dual Hybrid Gradient Solver (PDLP)

This directory contains PDLP, a library for solving linear programming (LP) and quadratic programming (QP) problems using first-order methods.

The implementation is based on the Primal-Dual Hybrid Gradient (PDHG) algorithm, which is preprocessed with scaling and optional presolving to improve performance and numerical stability. See also Mathematical background for PDLP.

Core C++ libraries:

• primal_dual_hybrid_gradient.h: The main entry point PrimalDualHybridGradient() for the solver, which takes a QuadraticProgram and solver parameters.
• quadratic_program.h: Defines the QuadraticProgram struct to represent the optimization problem, including objective vectors, constraint matrices, and bounds.
• quadratic_program_io.h: Provides utilities to read quadratic programs from various file formats, including MPS and MPModelProto.
• sharded_quadratic_program.h and sharder.h: These provide the infrastructure for sharding problem data and performing parallel computations.
• scheduler.h: A thread scheduling interface that supports multiple backends (e.g. Eigen's thread pools).
• iteration_stats.h and termination.h: Contain logic for computing convergence and infeasibility statistics and checking termination criteria.

Configuration and Logging

• solvers.proto: Defines the PrimalDualHybridGradientParams message for configuring the solver, including termination criteria, algorithmic choices like restart strategies, and linesearch rules.
• solve_log.proto: Defines messages for logging the solver's progress and final result, such as IterationStats and SolveLog.

Wrappers and Samples

• python/: Contains the pybind11 wrapper to expose the C++ library to Python, along with its build definitions and tests.
• samples/: This directory provides example usage of the library.