LBFGS++

‍UPDATE on 2020-03-06: LBFGS++ now includes a new L-BFGS-B solver for box-constrained optimization problems. Check the example below for its usage.

LBFGS++ is a header-only C++ library that implements the Limited-memory BFGS algorithm (L-BFGS) for unconstrained minimization problems, and a modified version of the L-BFGS-B algorithm for box-constrained ones.

The code for the L-BFGS solver is derived and modified from the libLBFGS library developed by Naoaki Okazaki.

LBFGS++ is implemented as a header-only C++ library, whose only dependency, Eigen, is also header-only.

A Quick Example

To use LBFGS++, one needs to first define a functor to represent the multivariate function to be minimized. It should return the objective function value on a vector x and overwrite the vector grad with the gradient evaluated on x. For example we could define the Rosenbrock function in the following way:

#include <Eigen/Core>
#include <iostream>
#include <LBFGS.h>
 
using Eigen::VectorXd;
using namespace LBFGSpp;
 
class Rosenbrock
{
private:
    int n;
public:
    Rosenbrock(int n_) : n(n_) {}
    double operator()(const VectorXd& x, VectorXd& grad)
    {
        double fx = 0.0;
        for(int i = 0; i < n; i += 2)
        {
            double t1 = 1.0 - x[i];
            double t2 = 10 * (x[i + 1] - x[i] * x[i]);
            grad[i + 1] = 20 * t2;
            grad[i]     = -2.0 * (x[i] * grad[i + 1] + t1);
            fx += t1 * t1 + t2 * t2;
        }
        return fx;
    }
};

Then we just need to set up parameters, create solver object, provide initial guess, and then run the minimization function.

int main()
{
    const int n = 10;
    // Set up parameters
    LBFGSParam<double> param;
    param.epsilon = 1e-6;
    param.max_iterations = 100;
 
    // Create solver and function object
    LBFGSSolver<double> solver(param);
    Rosenbrock fun(n);
 
    // Initial guess
    VectorXd x = VectorXd::Zero(n);
    // x will be overwritten to be the best point found
    double fx;
    int niter = solver.minimize(fun, x, fx);
 
    std::cout << niter << " iterations" << std::endl;
    std::cout << "x = \n" << x.transpose() << std::endl;
    std::cout << "f(x) = " << fx << std::endl;
 
    return 0;
}

The example can then be compiled and run.

$ g++ -I/path/to/eigen -I/path/to/lbfgspp/include -O2 example.cpp
$ ./a.out
23 iterations
x =
1 1 1 1 1 1 1 1 1 1
f(x) = 1.87948e-19

You can also use a different line search algorithm by providing a second template parameter to LBFGSSolver. For example, the code below illustrates the bracketing line search algorithm (contributed by @DirkToewe).

int main()
{
    const int n = 10;
    // Set up parameters
    LBFGSParam<double> param;
    param.epsilon = 1e-6;
    param.max_iterations = 100;
 
    // Create solver and function object
    LBFGSSolver<double, LineSearchBracketing> solver(param);
    Rosenbrock fun(n);
 
    // Initial guess
    VectorXd x = VectorXd::Zero(n);
    // x will be overwritten to be the best point found
    double fx;
    int niter = solver.minimize(fun, x, fx);
 
    std::cout << niter << " iterations" << std::endl;
    std::cout << "x = \n" << x.transpose() << std::endl;
    std::cout << "f(x) = " << fx << std::endl;
 
    return 0;
}

Box-constrained Problem

If the parameters to be optimized have simple bounds, then the L-BFGS-**B** solver class LBFGSBSolver can be used. The code is very similar to that of LBFGSSolver. Below is the same Rosenbrock example, but we require that all variables should be between 2 and 4.

#include <Eigen/Core>
#include <iostream>
#include <LBFGSB.h>  // Note the different header file
 
using Eigen::VectorXd;
using namespace LBFGSpp;
 
class Rosenbrock
{
private:
    int n;
public:
    Rosenbrock(int n_) : n(n_) {}
    double operator()(const VectorXd& x, VectorXd& grad)
    {
        double fx = 0.0;
        for(int i = 0; i < n; i += 2)
        {
            double t1 = 1.0 - x[i];
            double t2 = 10 * (x[i + 1] - x[i] * x[i]);
            grad[i + 1] = 20 * t2;
            grad[i]     = -2.0 * (x[i] * grad[i + 1] + t1);
            fx += t1 * t1 + t2 * t2;
        }
        return fx;
    }
};
 
int main()
{
    const int n = 10;
    // Set up parameters
    LBFGSBParam<double> param;  // New parameter class
    param.epsilon = 1e-6;
    param.max_iterations = 100;
 
    // Create solver and function object
    LBFGSBSolver<double> solver(param);  // New solver class
    Rosenbrock fun(n);
 
    // Bounds
    VectorXd lb = VectorXd::Constant(n, 2.0);
    VectorXd ub = VectorXd::Constant(n, 4.0);
 
    // Initial guess
    VectorXd x = VectorXd::Constant(n, 3.0);
 
    // x will be overwritten to be the best point found
    double fx;
    int niter = solver.minimize(fun, x, fx, lb, ub);
 
    std::cout << niter << " iterations" << std::endl;
    std::cout << "x = \n" << x.transpose() << std::endl;
    std::cout << "f(x) = " << fx << std::endl;
 
    return 0;
}

Note that we also allow infinite values for the lower and upper bounds. In such cases one can define ub[i] = std::numeric_limits<double>::infinity(), for example.

Documentation

The API reference page contains the documentation of LBFGS++ generated by Doxygen.

License

LBFGS++ is an open source project under the MIT license.