Numerical continuation in Python


Keywords
mathematics, numerical, continuation, methods, differential, equations, pde, python
Licenses
GPL-3.0+/OML
Install
pip install pacopy==0.2.8

Documentation

pacopy

PyPi Version PyPI pyversions GitHub stars PyPi downloads

Discord

pacopy provides various algorithms of numerical parameter continuation for ODEs and PDEs in Python.

pacopy is backend-agnostic, so it doesn't matter if your problem is formulated with NumPy, SciPy, FEniCS, pyfvm, or any other Python package. The only thing the user must provide is a class with some simple methods, e.g., a function evaluation f(u, lmbda), a Jacobian a solver jacobian_solver(u, lmbda, rhs) etc.

Install with

pip install pacopy

To get started, take a look at the examples below.

Some pacopy documentation is available here.

Examples

Basic scalar example

Let's start off with a problem where the solution space is scalar. We try to solve sin(x) - lambda for different values of lambda, starting at 0.

import math
import matplotlib.pyplot as plt
import pacopy


class SimpleScalarProblem:
    def inner(self, a, b):
        """The inner product in the problem domain. For scalars, this is just a
        multiplication.
        """
        return a * b

    def norm2_r(self, a):
        """The norm in the range space; used to determine if a solution has been found."""
        return a**2

    def f(self, u, lmbda):
        """The evaluation of the function to be solved"""
        return math.sin(u) - lmbda

    def df_dlmbda(self, u, lmbda):
        """The function's derivative with respect to the parameter. Used in Euler-Newton
        continuation.
        """
        return -1.0

    def jacobian_solver(self, u, lmbda, rhs):
        """A solver for the Jacobian problem. For scalars, this is just a division."""
        return rhs / math.cos(u)


problem = SimpleScalarProblem()

lmbda_list = []
values_list = []


def callback(k, lmbda, sol):
    # Use the callback for plotting, writing data to files etc.
    lmbda_list.append(lmbda)
    values_list.append(sol)


pacopy.euler_newton(
    problem,
    u0=0.0,
    lmbda0=0.0,
    callback=callback,
    max_steps=20,
    newton_tol=1.0e-10,
    verbose=False,
)

# plot solution
plt.plot(values_list, lmbda_list, ".-")
plt.xlabel("$u_1$")
plt.ylabel("$\\lambda$")
plt.show()

Simple 2D problem

A similarly simple example with two unknowns and a parameter. The inner product and Jacobian solver are getting more interesting.

import matplotlib.pyplot as plt
import numpy as np
import pacopy
import scipy
from mpl_toolkits.mplot3d import Axes3D


class SimpleProblem2D:
    def inner(self, a, b):
        return np.dot(a, b)

    def norm2_r(self, a):
        return np.dot(a, a)

    def f(self, u, lmbda):
        return np.array(
            [
                np.sin(u[0]) - lmbda - u[1] ** 2,
                np.cos(u[1]) * u[1] - lmbda,
            ]
        )

    def df_dlmbda(self, u, lmbda):
        return -np.ones_like(u)

    def jacobian_solver(self, u, lmbda, rhs):
        A = np.array(
            [
                [np.cos(u[0]), -2 * u[1]],
                [0.0, np.cos(u[1]) - np.sin(u[1]) * u[1]],
            ]
        )
        return np.linalg.solve(A, rhs)


problem = SimpleProblem2D()
# Initial guess and initial parameter value
u0 = np.zeros(2)
lmbda0 = 0.0

# Init lists
lmbda_list = []
values_list = []


def callback(k, lmbda, sol):
    lmbda_list.append(lmbda)
    values_list.append(sol)


pacopy.euler_newton(problem, u0, lmbda0, callback, max_steps=50, newton_tol=1.0e-10)

# plot solution
fig = plt.figure()
ax = fig.add_subplot(111, projection="3d")
ax.plot(*np.array(values_list).T, lmbda_list, ".-")
ax.set_xlabel("$u_1$")
ax.set_ylabel("$u_2$")
ax.set_zlabel("$\\lambda$")
# plt.show()
plt.savefig("simple2d.svg", transparent=True, bbox_inches="tight")
plt.close()

Bratu

Let's deal with an actual PDE, the classical Bratu problem in 1D with Dirichlet boundary conditions. Now, the solution space isn't scalar, but a vector of length n (the values of the solution at given points on the 1D interval). We now need numpy and scipy, the inner product and Jacobian solver are more complicated.

import matplotlib.pyplot as plt
import numpy as np
import pacopy
import scipy.sparse
import scipy.sparse.linalg


# This is the classical finite-difference approximation
class Bratu1d:
    def __init__(self, n):
        self.n = n
        h = 1.0 / (self.n - 1)

        self.H = np.full(self.n, h)
        self.H[0] = h / 2
        self.H[-1] = h / 2

        self.A = (
            scipy.sparse.diags([-1.0, 2.0, -1.0], [-1, 0, 1], shape=(self.n, self.n))
            / h**2
        )

    def inner(self, a, b):
        """The inner product of the problem. Can be np.dot(a, b), but factoring in
        the mesh width stays true to the PDE.
        """
        return np.dot(a, self.H * b)

    def norm2_r(self, a):
        """The norm in the range space; used to determine if a solution has been found."""
        return np.dot(a, a)

    def f(self, u, lmbda):
        """The evaluation of the function to be solved"""
        out = self.A.dot(u) - lmbda * np.exp(u)
        out[0] = u[0]
        out[-1] = u[-1]
        return out

    def df_dlmbda(self, u, lmbda):
        """The function's derivative with respect to the parameter. Used in Euler-Newton
        continuation.
        """
        out = -np.exp(u)
        out[0] = 0.0
        out[-1] = 0.0
        return out

    def jacobian_solver(self, u, lmbda, rhs):
        """A solver for the Jacobian problem."""
        M = self.A.copy()
        d = M.diagonal().copy()
        d -= lmbda * np.exp(u)
        M.setdiag(d)
        # Dirichlet conditions
        M.data[0][self.n - 2] = 0.0
        M.data[1][0] = 1.0
        M.data[1][self.n - 1] = 1.0
        M.data[2][1] = 0.0
        return scipy.sparse.linalg.spsolve(M.tocsr(), rhs)


problem = Bratu1d(51)
# Initial guess and parameter value
u0 = np.zeros(problem.n)
lmbda0 = 0.0

lmbda_list = []
values_list = []


def callback(k, lmbda, sol):
    # Use the callback for plotting, writing data to files etc.
    fig = plt.figure()
    ax1 = fig.add_subplot(111)
    ax1.set_xlabel("$\\lambda$")
    ax1.set_ylabel("$||u||_2$")
    ax1.grid()

    lmbda_list.append(lmbda)
    # use the norm of the currentsolution for plotting on the y-axis
    values_list.append(np.sqrt(problem.inner(sol, sol)))

    ax1.plot(lmbda_list, values_list, "-x", color="C0")
    ax1.set_xlim(0.0, 4.0)
    ax1.set_ylim(0.0, 6.0)
    plt.close()


# Natural parameter continuation
# pacopy.natural(problem, u0, lmbda0, callback, max_steps=100)

pacopy.euler_newton(
    problem, u0, lmbda0, callback, max_steps=500, max_num_retries=10, newton_tol=1.0e-10
)

Ginzburg–Landau

out.mp4

The Ginzburg-Landau equations model the behavior of extreme type-II superconductors under a magnetic field. The above example (to be found in full detail here) shows parameter continuation in the strength of the magnetic field. The plot on the right-hand side shows the complex-valued solution using cplot.