def test_add() -> None:
# Generate a few random points to use as test values.
np.random.seed(6178)
x_ = -1 + 2.0 * np.random.rand(100)
# A random number to use as an arbitrary additive constant.
# alpha = -0.194751428283640 + 0.079814485412665j
# Check operation in the face of empty arguments.
f1 = Function()
g1 = Function(lambda x: x)
assert len(f1 + f1) == 0
assert len(f1 + g1) == 0
assert len(g1 + f1) == 0
# Check addition of two function objects.
def f_op2(x):
return np.zeros(len(x))
f2 = Function(f_op2)
assert check_add_function_to_function(f2, f_op2, f2, f_op2, x_)
def f_op3(x):
return np.exp(x) - 1.0
f3 = Function(f_op3)
def g_op3(x):
return 1.0 / (1.0 + x**2)
g3 = Function(g_op3)
assert check_add_function_to_function(f3, f_op3, g3, g_op3, x_)
def g_op4(x):
return np.cos(1e4 * x)
g4 = Function(g_op4)
assert check_add_function_to_function(f3, f_op3, g4, g_op4, x_)
def g_op5(t):
return np.sinh(t * np.exp(2.0 * np.pi * 1.0j / 6.0))
g5 = Function(g_op5)
assert check_add_function_to_function(f3, f_op3, g5, g_op5, x_)
# Check that direct construction and PLUS give comparable results.
tol = 10 * np.spacing(1)
f = Function(lambda x: x)
g = Function(lambda x: np.cos(x) - 1.0)
h1 = f + g
h2 = Function(lambda x: x + np.cos(x) - 1.0)
# TODO: Improve the constructor so that the following passes:
assert np.linalg.norm(h1.coefficients[0] - h2.coefficients[0],
np.inf) < tol
def test_sub() -> None:
# Generate a few random points to use as test values.
x_ = -1.0 + 2.0 * np.random.rand(100)
# A random number to use as an arbitrary additive constant.
# alpha = np.random.randn() + 1.0j * np.random.randn()
# Check operation in the face of empty arguments.
f = Function()
g = Function(lambda x: x)
assert len(f - f) == 0
assert len(f - g) == 0
assert len(g - f) == 0
# Check subtraction of two function objects.
def f_op1(x):
return np.zeros(len(x))
f1 = Function(f_op1)
assert check_sub_function_and_function(f1, f_op1, f1, f_op1, x_)
def f_op2(x):
return np.exp(x) - 1
f2 = Function(f_op2)
def g_op1(x):
return 1.0 / (1 + x**2)
g1 = Function(g_op1)
assert check_sub_function_and_function(f2, f_op2, g1, g_op1, x_)
def g_op2(x):
return np.cos(1e4 * x)
g2 = Function(g_op2)
assert check_sub_function_and_function(f2, f_op2, g2, g_op2, x_)
def g_op3(t):
return np.sinh(t * np.exp(2.0 * np.pi * 1.0j / 6.0))
g3 = Function(g_op3)
assert check_sub_function_and_function(f2, f_op2, g3, g_op3, x_)
# Check that direct construction and the binary minus op give comparable results.
tol = 10.0 * np.spacing(1)
f = Function(lambda x: x)
g = Function(lambda x: np.cos(x) - 1)
h1 = f - g
h2 = Function(lambda x: x - (np.cos(x) - 1))
h3 = h1 - h2
assert np.linalg.norm(h3.coefficients, np.inf) < tol
def test_abs() -> None:
x_ = -1 + 2.0 * np.random.rand(100)
# Test a positive function:
def F1(x): # noqa
return np.sin(x) + 2.0
f = Function(lambda x: F1(x))
h = f.abs()
assert np.linalg.norm(h(x_) - f(x_), np.inf) < 10 * np.spacing(1)
# Test a negative function:
f2 = Function(lambda x: -F1(x))
h = f2.abs()
assert np.linalg.norm(h(x_) + f2(x_), np.inf) < 10 * np.spacing(1)
# Test a complex-valued function:
def F2(x_): # noqa
return np.exp(1.0j * np.pi * x_)
f = Function(lambda x: F2(x))
h = f.abs()
assert np.linalg.norm(h(x_) - 1.0, np.inf) < 1e2 * np.spacing(1)
def test_addition() -> None:
"""Test that it can sum a list of fractions."""
f1 = Function(lambda x: x)
g1 = Function(lambda x: -x)
h1 = f1 + g1
assert h1.iszero()
f2 = Function(lambda x: np.exp(-10 * x**2))
g2 = Function(lambda x: np.sin(2.7181828 * np.pi * x))
h2 = f2 + g2
xx = np.linspace(-1, 1, 201 + np.random.randint(0, 10))
error = f2(xx) + g2(xx) - h2(xx)
assert np.linalg.norm(error, np.inf) < 100 * 1e-16
def test_construction_discrete_data() -> None:
"""Test Constuction with discrete x and y data."""
x = np.linspace(0, 10, 101)
y = np.linspace(0, 10, 101)
func = Function(xdata=x, ydata=y)
assert np.allclose(func.domain, np.array([0.0, 10.0]), atol=1e-15)
assert len(func) == 2
x = np.random.rand(10)
y = np.random.rand(10)
func = Function(xdata=x, ydata=y)
error = func(x) - y
assert np.linalg.norm(error, np.inf) < 100 * 1.0e-6
def test_rmul() -> None:
"""Test the multiplication of f, specified by f_op, by a scalar alpha Generate a few random points to use as test
values.
:return:
"""
x_ = -1 + 2.0 * np.random.rand(100)
# Random numbers to use as arbitrary multiplicative constants.
alpha = -0.213251928283644 + 0.053493485412265j
# Check multiplication by scalars.
def f_op(x):
return np.sin(x)
f = Function(f_op)
g1 = f * alpha
g2 = alpha * f
assert g1 == g2
def g_exact(x):
return f_op(x) * alpha
assert np.linalg.norm(g1(x_) - g_exact(x_),
np.inf) < 10 * np.max(g1.vscale() * np.spacing(1))
def test_radd() -> None:
# Generate a few random points to use as test values.
np.random.seed(6178)
x_ = -1 + 2.0 * np.random.rand(100)
# A random number to use as an arbitrary additive constant.
alpha = -0.184752428910640 + 0.079812805462665j
# Check addition with scalars.
def f_op(x):
return np.sin(x)
f = Function(f_op)
# Test the addition of f, specified by f_op, to a scalar using
# a grid of points in [-1 1] for testing samples.
g1 = f + alpha
g2 = alpha + f
assert g1 == g2
def g_exact(x):
return f_op(x) + alpha
tol = 10 * g1.vscale() * np.spacing(1)
assert np.linalg.norm(g1(x_) - g_exact(x_), np.inf) <= tol
def test_construction_adaptive() -> None:
"""Test the construction process."""
f = Function()
assert len(f) == 0
assert f is not None
assert isinstance(f, Function)
# Set the tolerance:
tol = 100 * np.spacing(1)
# Test on a scalar-valued function:
def sin(t):
return np.sin(t)
g = Function(sin)
values = g.coefficients_to_values()
assert np.linalg.norm(sin(g.points) - values, np.inf) < tol
assert np.abs(g.vscale() - np.sin(1.0)) < tol
assert g.resolved
coeffs = np.array([0.0, 1.0])
result = Function(coefficients=coeffs)
assert np.all(result.coefficients == coeffs)
result = Function(fun=lambda x: x)
assert np.all(result.coefficients == np.array([0.0, 1.0]))
xx = np.linspace(-1, 1, 201 + np.random.randint(0, 10))
def f_true(x):
return np.exp(-10 * x**2)
f = Function(fun=f_true)
assert np.linalg.norm(f(xx) - f_true(xx), np.inf) < 100 * 1e-16
xx = np.linspace(-1, 1, 201 + np.random.randint(0, 10))
def g_true(x):
return np.sin(4 * np.pi * x)
g = Function(g_true)
assert np.linalg.norm(g(xx) - g_true(xx), np.inf) < 100 * 1e-16
def test_construction_fixed_length() -> None:
"""Test construction when length is given
:return:
"""
def fun(x):
return x
func = Function(fun, length=0)
assert len(func) == 0
func = Function(lambda x: x, length=2)
assert len(func) == 2
func = Function(lambda x: np.abs(x), lengths=[2, 2], domain=[-1, 0, 1])
assert len(func.pieces) == 2
assert len(func.pieces[0]) == 2
assert len(func.pieces[1]) == 2
func = Function(fun, length=201)
assert len(func) == 201
xx = np.linspace(-1, 1, 201 + np.random.randint(0, 10))
assert np.allclose(func(xx), xx, atol=1e-15)
assert np.allclose(func.points, func.values, atol=1e-15)
def spotcheck_min(fun_op: Callable, exact_min: complex) -> float:
# Spot-check the results for a given function.
f = Function(fun_op)
y = f.min()
x = f.argmin()
fx = fun_op(x)
result = (np.abs(y - exact_min) < 1.0e2 * f.vscale() * np.spacing(1)) and \
(np.abs(fx - exact_min) < 1.0e2 * f.vscale() * np.spacing(1))
return result
def polyfit_global(
x: np.ndarray,
y: np.ndarray,
degree: int = 1,
domain: Union[None, Sequence[float], np.ndarray] = None) -> Function:
"""Degree n least squares polynomial approximation of data y taken on points x in a domain.
:param x: x-values, np array
:param y: y-values, i.e., data values, np array
:param degree: degree of approximation, an integer
:param domain: domain of approximation
:return:
"""
if domain is None:
domain = [np.min(x), np.max(x)]
domain = 1.0 * np.array(domain)
a = domain[0]
b = domain[-1]
assert len(x) == len(
y), f"len(x) = {len(x)}, while len(y) = {len(y)}, these must be equal"
assert degree == int(
degree), f'degree = {degree}, degree must be an integer'
n = int(degree)
# map points to [-1, 1]
m = len(x)
x_normalized = 2.0 * (x - a) / (b - a) - 1.0
# construct the Chebyshev-Vandermonde matrix:
Tx = np.zeros((m, n + 1))
Tx[:, 0] = np.ones(m)
if n > 0:
Tx[:, 1] = x_normalized
for k in range(1, n):
Tx[:, k + 1] = 2.0 * x_normalized * Tx[:, k] - Tx[:, k - 1]
# c, residuals, rank, singular_values = polyfit_jit(x, y, n, a, b)
c, residuals, rank, singular_values = np.linalg.lstsq(Tx, y, rcond=None)
# Make a function:
return Function(coefficients=c, domain=domain)
def spotcheck_max(fun_op: Callable, exact_max: complex) -> float:
"""Spot-check the results for a given function."""
f = Function(fun_op)
y = f.max()
x = f.argmax()
fx = fun_op(x)
# [TODO]: Try to get this tolerance down:
result = (np.all(
np.abs(y - exact_max) < 1.0e2 * f.vscale() * np.spacing(1))) and (
np.all(
np.abs(fx - exact_max) < 1.0e2 * f.vscale() * np.spacing(1)))
return result
def polyfit(
x: np.ndarray,
y: np.ndarray,
degree: Union[int, Sequence[int]] = 1,
domain: Union[None, np.ndarray, Sequence[float]] = None) -> Function:
"""Least squares polynomial fitting to discrete data with piecewise domain splitting handled.
:param x:
:param y:
:param degree: an array or a double len(degree) = len(domain) - 1
:param domain: an array to specify where the breakpionts are
:return:
"""
if domain is None or len(domain) == 2:
assert isinstance(degree, int)
return polyfit_global(x, y, degree, domain)
if len(domain) > 2:
n_pieces = len(domain) - 1
if isinstance(degree, Sequence): # A list of degrees is passed
assert n_pieces == len(
degree), 'must specify degree for each domain'
degrees = degree
else: # The degree passed is just an integer
degrees = n_pieces * [int(degree)]
else:
raise AssertionError(
f'domain = {domain}, domain must be for the form [a, b]')
all_coefficients = n_pieces * [None]
for j in range(n_pieces):
a, b = domain[j], domain[j + 1]
idx = (a <= x) & (x <= b)
if np.sum(idx) == 0:
c = np.array([])
else:
xj, yj = x[idx], y[idx]
f = polyfit_global(xj, yj, degree=degrees[j], domain=[a, b])
c = f.coefficients[0].copy()
all_coefficients[j] = c
return Function(coefficients=all_coefficients, domain=domain)
def test_rsub() -> None:
# Generate a few random points to use as test values.
x_ = -1.0 + 2.0 * np.random.rand(100)
# A random number to use as an arbitrary additive constant.
alpha = np.random.randn() + 1.0j * np.random.randn()
def f_op(x):
return np.sin(x)
f = Function(f_op)
# Test the subtraction of f, to and from a scalar using a grid of points
g1 = f - alpha
g2 = alpha - f
assert g1 == g2
def g_exact(x):
return f_op(x) - alpha
# [TODO] can we bring this tolerance down?
tol = 1.0e2 * g1.vscale() * np.spacing(1)
assert np.linalg.norm(g1(x_) - g_exact(x_), np.inf) <= tol
def test_roots() -> None:
def func(x):
return (x + 1) * 50
f = Function(lambda x: special.j0(func(x)))
r = func(f.roots())
exact = np.array([
2.40482555769577276862163, 5.52007811028631064959660,
8.65372791291101221695437, 11.7915344390142816137431,
14.9309177084877859477626, 18.0710639679109225431479,
21.2116366298792589590784, 24.3524715307493027370579,
27.4934791320402547958773, 30.6346064684319751175496,
33.7758202135735686842385, 36.9170983536640439797695,
40.0584257646282392947993, 43.1997917131767303575241,
46.3411883716618140186858, 49.4826098973978171736028,
52.6240518411149960292513, 55.7655107550199793116835,
58.9069839260809421328344, 62.0484691902271698828525,
65.1899648002068604406360, 68.3314693298567982709923,
71.4729816035937328250631, 74.6145006437018378838205,
77.7560256303880550377394, 80.8975558711376278637723,
84.0390907769381901578795, 87.1806298436411536512617,
90.3221726372104800557177, 93.4637187819447741711905,
96.6052679509962687781216, 99.7468198586805964702799
])
assert np.linalg.norm(r - exact, np.inf) < 1.0e1 * len(f) * np.spacing(1)
k = 500
f = Function(lambda x: np.sin(np.pi * k * x))
r = f.roots()
assert np.linalg.norm(r - (1.0 * np.r_[-k:k + 1]) / k,
np.inf) < 1e1 * len(f) * np.spacing(1)
# Test a perturbed polynomial:
f = Function(lambda x: (x - .1) * (x + .9) * x * (x - .9) + 1e-14 * x**5)
r = f.roots()
assert len(r) == 4
assert np.linalg.norm(f(r), np.inf) < 1e2 * len(f) * np.spacing(1)
# Test a some simple polynomials:
f = Function(values=[-1.0, 1.0])
r = f.roots()
assert np.all(r == 0)
# f = testclass.make([1 0 1])
f = Function(values=[1.0, 0.0, 1.0])
r = f.roots()
assert len(r) == 2
assert np.linalg.norm(r, np.inf) < np.spacing(1)
# Test some complex roots:
f = Function(lambda x: 1 + 25 * x**2)
r = f.roots(complex_roots=True)
assert len(r) == 2
assert np.linalg.norm(r - np.r_[-1.0j, 1.0j] / 5.0,
np.inf) < 10 * np.spacing(1)
# [TODO] This is failing:
# f = Function(lambda x: (1 + 25*x**2)*np.exp(x))
# r = f.roots(complex_roots=True, prune=True)
# assert len(r), 2)
# assert np.linalg.norm( r - np.r_[1.0j, -1.0j]/5.0, np.inf) < 10*len(f)*np.spacing(1))
# [TODO] We get different number of roots
# f = Function(lambda x: np.sin(100*np.pi*x))
# r1 = f.roots(complex_roots=True, recurse=False)
# r2 = f.roots(complex_roots=True)
# assert len(r1), 201)
# assert len(r2), 213)
# Adding test for 'qz' flag:
f = Function(lambda x: 1e-10 * x**3 + x**2 - 1e-12)
r = f.roots(qz=True)
assert len(r) != 0
assert np.linalg.norm(f[r], np.inf) < 10 * np.spacing(1)
# Add a rootfinding test for low degree non-even functions:
f = Function(lambda x: (x - .5) * (x - 1.0 / 3.0))
r = f.roots(qz=True)
assert np.linalg.norm(f[r], np.inf) < np.spacing(1)
def test_mul() -> None:
# Generate a few random points to use as test values.
np.random.seed(6178)
x_ = -1 + 2.0 * np.random.rand(100)
# Random numbers to use as arbitrary multiplicative constants.
alpha = -0.194758928283640 + 0.075474485412665j
# Check operation in the face of empty arguments.
f1 = Function()
g1 = Function(lambda x: x)
assert len(f1 * f1) == 0
assert len(f1 * g1) == 0
assert len(g1 * f1) == 0
# Check multiplication by constant functions.
def f_op2(x):
return np.sin(x)
f2 = Function(f_op2)
def g_op2(x):
return np.zeros(len(x)) + alpha
g2 = Function(g_op2)
assert check_mul_function_by_function(f2, f_op2, g2, g_op2, x_, False)
# Spot-check multiplication of two function objects for a few test
# functions.
def f_op3(x):
return np.ones(len(x))
f3 = Function(f_op3)
assert check_mul_function_by_function(f3, f_op3, f3, f_op3, x_, False)
def f_op4(x):
return np.exp(x) - 1.0
f4 = Function(f_op4)
def g_op3(x):
return 1.0 / (1.0 + x**2)
g3 = Function(g_op3)
assert check_mul_function_by_function(f4, f_op4, g3, g_op3, x_, False)
# If f and g are real then so must be f * g
h = f4 * g3
assert h.isreal()
def g_op4(x):
return np.cos(1.0e4 * x)
g4 = Function(g_op4)
assert check_mul_function_by_function(f4, f_op4, g4, g_op4, x_, False)
def g_op5(t):
return np.sinh(t * np.exp(2.0 * np.pi * 1.0j / 6.0))
g5 = Function(g_op5)
assert check_mul_function_by_function(f4, f_op4, g5, g_op5, x_, False)
# Check specially handled cases, including some in which an adjustment for
# positivity is performed.
def f_op5(t):
return np.sinh(t * np.exp(2.0 * np.pi * 1.0j / 6.0))
f5 = Function(f_op5)
assert check_mul_function_by_function(f5, f_op5, f5, f_op5, x_, False)
def g_op6(t):
return np.conjugate(np.sinh(t * np.exp(2.0 * np.pi * 1.0j / 6.0)))
g6 = f5.conj()
assert check_mul_function_by_function(f5, f_op5, g6, g_op6, x_, True)
def f_op7(x):
return np.exp(x) - 1.0
f7 = Function(f_op7)
assert check_mul_function_by_function(f7, f_op7, f7, f_op7, x_, True)
# Check that multiplication and direct construction give similar results.
tol = 50 * np.spacing(1)
def g_op7(x):
return 1.0 / (1.0 + x**2)
g7 = Function(g_op7)
h1 = f7 * g7
h2 = Function(lambda x: f_op7(x) * g_op7(x))
h2 = h2.prolong(len(h1))
assert np.linalg.norm(h1.coefficients[0] - h2.coefficients[0],
np.inf) < tol
def test_poly() -> None:
f = Function()
assert len(f.poly()[0]) == 0
f = Function(lambda x: 1.0 + 0.0 * x)
assert np.linalg.norm(f.poly()[0] - np.ones(1),
np.inf) < f.vscale() * np.spacing(1)
f = Function(lambda x: 1 + x)
assert np.linalg.norm(f.poly()[0] - np.ones(2),
np.inf) < f.vscale() * np.spacing(1)
f = Function(lambda x: 1 + x + x**2)
assert np.linalg.norm(f.poly()[0] - np.ones(3),
np.inf) < f.vscale() * np.spacing(1)
f = Function(lambda x: 1 + x + x**2 + x**3)
assert np.linalg.norm(f.poly()[0] - np.ones(4),
np.inf) < f.vscale() * np.spacing(1)
f = Function(lambda x: 1 + x + x**2 + x**3 + x**4)
assert np.linalg.norm(f.poly()[0] - np.ones(5),
np.inf) < f.vscale() * np.spacing(1)
import matplotlib.pyplot as plt
import numpy as np
from numfun.function import Function
f = Function(lambda x: np.heaviside(x, 0), domain=[-1, 0, 1])
# Define a probability distribution
a = 0
b = 100
domain = [a, b]
x = Function(lambda x: x, domain=domain)
f = Function(lambda x: 2.0 * np.exp(-2.0 * x), domain=domain)
f.plot()
# What is the expected value and the variance:
E = (x * f).definite_integral()
V = (x**2 * f).definite_integral() - E**2
#%%
f = Function(lambda x: np.sin(x) + np.sin(5 * x**2), domain=[0, 10])
x = f.roots()
y = f(x)
plt.figure(figsize=(10, 8))
f.plot()
plt.plot(x, y, '.')
#%%
f = Function(fun=[
lambda x: np.exp(-1.0 / x**2), lambda x: 0 * x,
lambda x: np.exp(-1. / x**2)
def test_cumsum() -> None:
# Generate a few random points to use as test values.
x_ = 2 * np.random.rand(100) - 1
# Spot-check antiderivatives for a couple of functions. We verify that the
# function antiderivatives match the true ones up to a constant by checking
# that the standard deviation of the difference between the two on a large
# random grid is small. We also check that evaluate(f.cumsum(), -1) == 0 each
# time.
f = Function(lambda x: np.exp(x) - 1.0)
F = f.cumsum() # noqa
F_ex = lambda x: np.exp(x) - x # noqa
err = np.std(F[x_] - F_ex(x_))
tol = 20 * F.vscale() * np.spacing(1)
assert err < tol
assert np.abs(F[-1]) < tol
f = Function(lambda x: 1.0 / (1.0 + x**2))
F = f.cumsum() # noqa
F_ex = lambda x: np.arctan(x) # noqa
err = np.std(F[x_] - F_ex(x_))
tol = 10 * F.vscale() * np.spacing(1)
assert err < tol
assert np.abs(F[-1]) < tol
f = Function(lambda x: np.cos(1.0e4 * x))
F = f.cumsum() # noqa
F_ex = lambda x: np.sin(1.0e4 * x) / 1.0e4 # noqa
err = F[x_] - F_ex(x_)
tol = 10.0e4 * F.vscale() * np.spacing(1)
assert (np.std(err) < tol) and (np.abs(F[-1]) < tol)
z = np.exp(2 * np.pi * 1.0j / 6)
f = Function(lambda t: np.sinh(t * z))
F = f.cumsum() # noqa
F_ex = lambda t: np.cosh(t * z) / z # noqa
err = F[x_] - F_ex(x_)
tol = 10 * F.vscale() * np.spacing(1)
assert (np.std(err) < tol) and (np.abs(F[-1]) < tol)
# Check that applying cumsum() and direct construction of the antiderivative
# give the same results (up to a constant).
f = Function(lambda x: np.sin(4.0 * x)**2)
F = Function(lambda x: 0.5 * x - 0.0625 * np.sin(8 * x)) # noqa
G = f.cumsum() # noqa
err = G - F
tol = 10 * G.vscale() * np.spacing(1)
values = err.coefficients_to_values(err.coefficients)
assert (np.std(values) < tol) and (np.abs(G[-1]) < tol)
# Check that f.diff().cumsum() == f and that f.cumsum().diff() == f up to a
# constant.
f = Function(lambda x: x * (x - 1.0) * np.sin(x) + 1.0)
# def integral_f(x):
# return (-x ** 2 + x + 2) * np.cos(x) + x + (2 * x - 1) * np.sin(x)
# F = lambda x: integral_f(x) - integral_f(-1)
g = f.cumsum().diff()
err = f(x_) - g(x_)
tol = 10 * g.vscale() * np.spacing(1)
assert np.linalg.norm(err, np.inf) < 100 * tol
h = f.diff().cumsum()
err = f(x_) - h(x_)
tol = 10 * h.vscale() * np.spacing(1)
assert (np.std(err) < tol) and (np.abs(h[-1]) < tol)
def y_true(xx):
return a + b * xx + c * xx ** 2
def y_noisy(xx):
return y_true(xx) + w
y = y_noisy(x)
# % interpolation in equidistant points doesn't really
# work:
g = Function(lambda xx: barycentric_interpolation(xx, y, x), length=len(y), domain=[0, 5])
g.plot()
plt.plot(x, y, 'g.', x, y - w, 'k--', x, g(x), 'r.')
plt.grid(True)
plt.show()
# % So we try cubic splines:
# x = np.arange(10)
# y = np.sin(x)
cs = CubicSpline(x, y)
xs = np.arange(0.0, 5, 0.05)
ys = y_true(xs)
plt.figure(figsize=(10, 8))
plt.plot(x, y, 'o', label='obs data')
plt.plot(xs, ys, label='true data')
plt.plot(xs, cs(xs), label="Spline")
def test_sum() -> None:
# Spot-check integrals for a couple of functions.
f = Function(lambda x: np.exp(x) - 1.0)
assert np.abs(f.sum() -
0.350402387287603) < 10 * f.vscale() * np.spacing(1)
f = Function(lambda x: 1. / (1 + x**2))
assert np.abs(f.sum() - np.pi / 2.0) < 10 * f.vscale() * np.spacing(1)
f = Function(lambda x: np.cos(1e4 * x))
exact = -6.112287777765043e-05
assert np.abs(f.sum() -
exact) / np.abs(exact) < 1e6 * f.vscale() * np.spacing(1)
z = np.exp(2 * np.pi * 1.0j / 6.0)
f = Function(lambda t: np.sinh(t * z))
assert np.abs(f.sum()) < 10 * f.vscale() * np.spacing(1)
# Check a few basic properties.
a = 2.0
b = -1.0j
f = Function(lambda x: x * np.sin(x**2) - 1)
df = f.diff()
g = Function(lambda x: np.exp(-x**2))
dg = g.diff()
fg = f * g
gdf = g * df
fdg = f * dg
tol_f = 10 * f.vscale() * np.spacing(1)
tol_g = 10 * f.vscale() * np.spacing(1)
tol_df = 10 * df.vscale() * np.spacing(1)
tol_dg = 10 * dg.vscale() * np.spacing(1)
tol_fg = 10 * fg.vscale() * np.spacing(1)
tol_fdg = 10 * fdg.vscale() * np.spacing(1)
tol_gdf = 10 * gdf.vscale() * np.spacing(1)
# Linearity.
assert np.abs((a * f + b * g).sum() -
(a * f.sum() + b * g.sum())) < max(tol_f, tol_g)
# Integration-by-parts.
assert np.abs(fdg.sum() - (fg(1) - fg(-1) - gdf.sum())) < np.max(
np.r_[tol_fdg, tol_gdf, tol_fg])
# Fundamental Theorem of Calculus.
assert np.abs(df.sum() - (f(1) - f(-1))) < np.max(np.r_[tol_df, tol_f])
assert np.abs(dg.sum() - (g(1) - g(-1))) < np.max(np.r_[tol_dg, tol_g])
