jacobi 0.2.1

Features

• Robustly compute the generalised Jacobi matrix for an arbitrary real analytic mapping ℝⁿ → ℝⁱ¹ × … × ℝⁱⁿ

• Derivative is either computed to specified accuracy (to save computing time) or until maximum precision of function is reached

• Algorithm based on John D’Errico’s DERIVEST: works even with functions that have large round-off error

• Up to 1200x faster than numdifftools at equivalent precision

• Returns error estimates for derivatives

• Supports arbitrary auxiliary function arguments

• Perform statistical error propagation based on numerically computed jacobian

• Lightweight package, only depends on numpy

Planned features

• Compute the Hessian matrix numerically with the same algorithm

• Further generalize the calculation to support function arguments with shape (N, K), in that case compute the Jacobi matrix for each of the K vectors of length N

Examples

from matplotlib import pyplot as plt
import numpy as np
from jacobi import jacobi


# function of one variable with auxiliary argument; returns a vector
def f(dx, x):
    y = x + dx
    return np.sin(y) / y


x = np.linspace(-10, 10, 1000)
fx = f(0, x)
fdx, fdex = jacobi(f, 0, x)

plt.plot(x, fx, label="f(x) = sin(x) / x")
plt.plot(x, fdx, ls="--", label="f'(x)")
plt.legend()

from jacobi import propagate
import numpy as np


def f(x):
    a = 1.5
    b = 3.1
    return a * np.exp(-x ** 2) + b

x = [1.0, 2.0]
xcov = [[1.1, 0.1], [0.1, 2.3]]
y, ycov = propagate(f, x, xcov)

Comparison to numdifftools

Speed

Jacobi makes better use of vectorised computation than numdifftools. Smaller run-time is better (and ratio > 1).

Precision

The machine precision is indicated by the dashed line.