def test_JointEigenDistribution():
A = Matrix([[Normal('A00', 0, 1), Normal('A01', 1, 1)],
[Beta('A10', 1, 1), Beta('A11', 1, 1)]])
JointEigenDistribution(A) == \
JointDistributionHandmade(-sqrt(A[0, 0]**2 - 2*A[0, 0]*A[1, 1] + 4*A[0, 1]*A[1, 0] + A[1, 1]**2)/2 +
A[0, 0]/2 + A[1, 1]/2, sqrt(A[0, 0]**2 - 2*A[0, 0]*A[1, 1] + 4*A[0, 1]*A[1, 0] + A[1, 1]**2)/2 + A[0, 0]/2 + A[1, 1]/2)
raises(ValueError, lambda: JointEigenDistribution(Matrix([[1, 0], [2, 1]])))
def test_sample_numpy():
distribs_numpy = [
Beta("B", 1, 1),
Normal("N", 0, 1),
Gamma("G", 2, 7),
Exponential("E", 2),
LogNormal("LN", 0, 1),
Pareto("P", 1, 1),
ChiSquared("CS", 2),
Uniform("U", 0, 1)
]
size = 3
numpy = import_module('numpy')
if not numpy:
skip('Numpy is not installed. Abort tests for _sample_numpy.')
else:
for X in distribs_numpy:
samps = sample(X, size=size, library='numpy')
for sam in samps:
assert sam in X.pspace.domain.set
raises(NotImplementedError,
lambda: sample(Chi("C", 1), library='numpy'))
raises(
NotImplementedError,
lambda: Chi("C", 1).pspace.distribution.sample(library='tensorflow'))
def test_beta():
a, b = symbols('alpha beta', positive=True)
B = Beta('x', a, b)
assert pspace(B).domain.set == Interval(0, 1)
assert characteristic_function(B)(x) == hyper((a,), (a + b,), I*x)
assert density(B)(x) == x**(a - 1)*(1 - x)**(b - 1)/beta(a, b)
assert simplify(E(B)) == a / (a + b)
assert simplify(variance(B)) == a*b / (a**3 + 3*a**2*b + a**2 + 3*a*b**2 + 2*a*b + b**3 + b**2)
# Full symbolic solution is too much, test with numeric version
a, b = 1, 2
B = Beta('x', a, b)
assert expand_func(E(B)) == a / S(a + b)
assert expand_func(variance(B)) == (a*b) / S((a + b)**2 * (a + b + 1))
def test_precomputed_characteristic_functions():
import mpmath
def test_cf(dist, support_lower_limit, support_upper_limit):
pdf = density(dist)
t = Symbol('t')
x = Symbol('x')
# first function is the hardcoded CF of the distribution
cf1 = lambdify([t], characteristic_function(dist)(t), 'mpmath')
# second function is the Fourier transform of the density function
f = lambdify([x, t], pdf(x)*exp(I*x*t), 'mpmath')
cf2 = lambda t: mpmath.quad(lambda x: f(x, t), [support_lower_limit, support_upper_limit], maxdegree=10)
# compare the two functions at various points
for test_point in [2, 5, 8, 11]:
n1 = cf1(test_point)
n2 = cf2(test_point)
assert abs(re(n1) - re(n2)) < 1e-12
assert abs(im(n1) - im(n2)) < 1e-12
test_cf(Beta('b', 1, 2), 0, 1)
test_cf(Chi('c', 3), 0, mpmath.inf)
test_cf(ChiSquared('c', 2), 0, mpmath.inf)
test_cf(Exponential('e', 6), 0, mpmath.inf)
test_cf(Logistic('l', 1, 2), -mpmath.inf, mpmath.inf)
test_cf(Normal('n', -1, 5), -mpmath.inf, mpmath.inf)
test_cf(RaisedCosine('r', 3, 1), 2, 4)
test_cf(Rayleigh('r', 0.5), 0, mpmath.inf)
test_cf(Uniform('u', -1, 1), -1, 1)
test_cf(WignerSemicircle('w', 3), -3, 3)
def test_sample_pymc3():
distribs_pymc3 = [
Beta("B", 1, 1),
Cauchy("C", 1, 1),
Normal("N", 0, 1),
Gamma("G", 2, 7),
GaussianInverse("GI", 1, 1),
Exponential("E", 2),
LogNormal("LN", 0, 1),
Pareto("P", 1, 1),
ChiSquared("CS", 2),
Uniform("U", 0, 1)
]
size = 3
pymc3 = import_module('pymc3')
if not pymc3:
skip('PyMC3 is not installed. Abort tests for _sample_pymc3.')
else:
with ignore_warnings(
UserWarning
): ### TODO: Restore tests once warnings are removed
for X in distribs_pymc3:
samps = next(sample(X, size=size, library='pymc3'))
for sam in samps:
assert sam in X.pspace.domain.set
raises(NotImplementedError,
lambda: next(sample(Chi("C", 1), library='pymc3')))
def test_Compound_Distribution():
X = Normal('X', 2, 4)
N = NormalDistribution(X, 4)
C = CompoundDistribution(N)
assert C.is_Continuous
assert C.set == Interval(-oo, oo)
assert C.pdf(x, evaluate=True).simplify() == exp(-x**2/64 + x/16 - S(1)/16)/(8*sqrt(pi))
assert not isinstance(CompoundDistribution(NormalDistribution(2, 3)),
CompoundDistribution)
M = MultivariateNormalDistribution([1, 2], [[2, 1], [1, 2]])
raises(NotImplementedError, lambda: CompoundDistribution(M))
X = Beta('X', 2, 4)
B = BernoulliDistribution(X, 1, 0)
C = CompoundDistribution(B)
assert C.is_Finite
assert C.set == {0, 1}
y = symbols('y', negative=False, integer=True)
assert C.pdf(y, evaluate=True) == Piecewise((S(1)/(30*beta(2, 4)), Eq(y, 0)),
(S(1)/(60*beta(2, 4)), Eq(y, 1)), (0, True))
k, t, z = symbols('k t z', positive=True, real=True)
G = Gamma('G', k, t)
X = PoissonDistribution(G)
C = CompoundDistribution(X)
assert C.is_Discrete
assert C.set == S.Naturals0
assert C.pdf(z, evaluate=True).simplify() == t**z*(t + 1)**(-k - z)*gamma(k \
+ z)/(gamma(k)*gamma(z + 1))
def test_density():
x = Symbol('x')
l = Symbol('l', positive=True)
rate = Beta(l, 2, 3)
X = Poisson(x, rate)
assert isinstance(pspace(X), ProductPSpace)
assert density(X, Eq(rate, rate.symbol)) == PoissonDistribution(l)
def test_sample_scipy():
distribs_scipy = [
Beta("B", 1, 1),
BetaPrime("BP", 1, 1),
Cauchy("C", 1, 1),
Chi("C", 1),
Normal("N", 0, 1),
Gamma("G", 2, 7),
GammaInverse("GI", 1, 1),
GaussianInverse("GUI", 1, 1),
Exponential("E", 2),
LogNormal("LN", 0, 1),
Pareto("P", 1, 1),
StudentT("S", 2),
ChiSquared("CS", 2),
Uniform("U", 0, 1)
]
size = 3
scipy = import_module('scipy')
if not scipy:
skip('Scipy is not installed. Abort tests for _sample_scipy.')
else:
for X in distribs_scipy:
samps = sample(X, size=size, library='scipy')
samps2 = sample(X, size=(2, 2), library='scipy')
for sam in samps:
assert sam in X.pspace.domain.set
for i in range(2):
for j in range(2):
assert samps2[i][j] in X.pspace.domain.set
def test_beta():
a, b = symbols('alpha beta', positive=True)
B = Beta('x', a, b)
assert pspace(B).domain.set == Interval(0, 1)
dens = density(B)
x = Symbol('x')
assert dens(x) == x**(a - 1)*(1 - x)**(b - 1) / beta(a, b)
# This is too slow
# assert E(B) == a / (a + b)
# assert variance(B) == (a*b) / ((a+b)**2 * (a+b+1))
# Full symbolic solution is too much, test with numeric version
a, b = 1, 2
B = Beta('x', a, b)
assert expand_func(E(B)) == a / S(a + b)
assert expand_func(variance(B)) == (a*b) / S((a + b)**2 * (a + b + 1))
def test_density():
x = Symbol('x')
l = Symbol('l', positive=True)
rate = Beta(l, 2, 3)
X = Poisson(x, rate)
assert isinstance(pspace(X), JointPSpace)
assert density(X, Eq(rate, rate.symbol)) == PoissonDistribution(l)
N1 = Normal('N1', 0, 1)
N2 = Normal('N2', N1, 2)
assert density(N2)(0).doit() == sqrt(10) / (10 * sqrt(pi))
assert simplify(density(N2, Eq(N1, 1))(x)) == \
sqrt(2)*exp(-(x - 1)**2/8)/(4*sqrt(pi))
def test_prefab_sampling():
N = Normal('X', 0, 1)
L = LogNormal('L', 0, 1)
E = Exponential('Ex', 1)
P = Pareto('P', 1, 3)
W = Weibull('W', 1, 1)
U = Uniform('U', 0, 1)
B = Beta('B', 2, 5)
G = Gamma('G', 1, 3)
variables = [N, L, E, P, W, U, B, G]
niter = 10
for var in variables:
for i in range(niter):
assert sample(var) in var.pspace.domain.set
def gen_samples(self, n, r, a=None, b=None):
if not a: a = self.a
if not b: b = self.b
samples = np.empty((0, r))
for _ in range(n):
p = Beta('p', a, b)
B = Binomial('B', 1, sym_sample(p))
sample = []
for _ in range(r):
# TODO: change to sample a complete vector at once
sample.append(sym_sample(B))
samples = np.append(samples, [sample], axis=0)
return samples
def test_prefab_sampling():
N = Normal(0, 1)
L = LogNormal(0, 1)
E = Exponential(1)
P = Pareto(1, 3)
W = Weibull(1, 1)
U = Uniform(0, 1)
B = Beta(2,5)
G = Gamma(1,3)
variables = [N,L,E,P,W,U,B,G]
niter = 10
for var in variables:
for i in xrange(niter):
assert Sample(var) in var.pspace.domain.set
def test_bernoulli_CompoundDist():
X = Beta('X', 1, 2)
Y = Bernoulli('Y', X)
assert density(Y).dict == {0: S(2)/3, 1: S(1)/3}
assert E(Y) == P(Eq(Y, 1)) == S(1)/3
assert variance(Y) == S(2)/9
assert cdf(Y) == {0: S(2)/3, 1: 1}
# test issue 8128
a = Bernoulli('a', S(1)/2)
b = Bernoulli('b', a)
assert density(b).dict == {0: S(1)/2, 1: S(1)/2}
assert P(b > 0.5) == S(1)/2
X = Uniform('X', 0, 1)
Y = Bernoulli('Y', X)
assert E(Y) == S(1)/2
assert P(Eq(Y, 1)) == E(Y)
def test_prefab_sampling():
scipy = import_module('scipy')
if not scipy:
skip('Scipy is not installed. Abort tests')
N = Normal('X', 0, 1)
L = LogNormal('L', 0, 1)
E = Exponential('Ex', 1)
P = Pareto('P', 1, 3)
W = Weibull('W', 1, 1)
U = Uniform('U', 0, 1)
B = Beta('B', 2, 5)
G = Gamma('G', 1, 3)
variables = [N, L, E, P, W, U, B, G]
niter = 10
size = 5
for var in variables:
for _ in range(niter):
assert sample(var) in var.pspace.domain.set
samps = sample(var, size=size)
for samp in samps:
assert samp in var.pspace.domain.set
def test_sample_scipy():
distribs_scipy = [
Beta("B", 1, 1),
BetaPrime("BP", 1, 1),
Cauchy("C", 1, 1),
Chi("C", 1),
Normal("N", 0, 1),
Gamma("G", 2, 7),
GammaInverse("GI", 1, 1),
GaussianInverse("GUI", 1, 1),
Exponential("E", 2),
LogNormal("LN", 0, 1),
Pareto("P", 1, 1),
StudentT("S", 2),
ChiSquared("CS", 2),
Uniform("U", 0, 1)
]
size = 3
numsamples = 5
scipy = import_module('scipy')
if not scipy:
skip('Scipy is not installed. Abort tests for _sample_scipy.')
else:
with ignore_warnings(
UserWarning
): ### TODO: Restore tests once warnings are removed
g_sample = list(
sample(Gamma("G", 2, 7), size=size, numsamples=numsamples))
assert len(g_sample) == numsamples
for X in distribs_scipy:
samps = next(sample(X, size=size, library='scipy'))
samps2 = next(sample(X, size=(2, 2), library='scipy'))
for sam in samps:
assert sam in X.pspace.domain.set
for i in range(2):
for j in range(2):
assert samps2[i][j] in X.pspace.domain.set
def test_moment_generating_function():
t = symbols('t', positive=True)
# Symbolic tests
a, b, c = symbols('a b c')
mgf = moment_generating_function(Beta('x', a, b))(t)
assert mgf == hyper((a, ), (a + b, ), t)
mgf = moment_generating_function(Chi('x', a))(t)
assert mgf == sqrt(2)*t*gamma(a/2 + S.Half)*\
hyper((a/2 + S.Half,), (Rational(3, 2),), t**2/2)/gamma(a/2) +\
hyper((a/2,), (S.Half,), t**2/2)
mgf = moment_generating_function(ChiSquared('x', a))(t)
assert mgf == (1 - 2 * t)**(-a / 2)
mgf = moment_generating_function(Erlang('x', a, b))(t)
assert mgf == (1 - t / b)**(-a)
mgf = moment_generating_function(ExGaussian("x", a, b, c))(t)
assert mgf == exp(a * t + b**2 * t**2 / 2) / (1 - t / c)
mgf = moment_generating_function(Exponential('x', a))(t)
assert mgf == a / (a - t)
mgf = moment_generating_function(Gamma('x', a, b))(t)
assert mgf == (-b * t + 1)**(-a)
mgf = moment_generating_function(Gumbel('x', a, b))(t)
assert mgf == exp(b * t) * gamma(-a * t + 1)
mgf = moment_generating_function(Gompertz('x', a, b))(t)
assert mgf == b * exp(b) * expint(t / a, b)
mgf = moment_generating_function(Laplace('x', a, b))(t)
assert mgf == exp(a * t) / (-b**2 * t**2 + 1)
mgf = moment_generating_function(Logistic('x', a, b))(t)
assert mgf == exp(a * t) * beta(-b * t + 1, b * t + 1)
mgf = moment_generating_function(Normal('x', a, b))(t)
assert mgf == exp(a * t + b**2 * t**2 / 2)
mgf = moment_generating_function(Pareto('x', a, b))(t)
assert mgf == b * (-a * t)**b * uppergamma(-b, -a * t)
mgf = moment_generating_function(QuadraticU('x', a, b))(t)
assert str(mgf) == (
"(3*(t*(-4*b + (a + b)**2) + 4)*exp(b*t) - "
"3*(t*(a**2 + 2*a*(b - 2) + b**2) + 4)*exp(a*t))/(t**2*(a - b)**3)")
mgf = moment_generating_function(RaisedCosine('x', a, b))(t)
assert mgf == pi**2 * exp(a * t) * sinh(b * t) / (b * t *
(b**2 * t**2 + pi**2))
mgf = moment_generating_function(Rayleigh('x', a))(t)
assert mgf == sqrt(2)*sqrt(pi)*a*t*(erf(sqrt(2)*a*t/2) + 1)\
*exp(a**2*t**2/2)/2 + 1
mgf = moment_generating_function(Triangular('x', a, b, c))(t)
assert str(mgf) == ("(-2*(-a + b)*exp(c*t) + 2*(-a + c)*exp(b*t) + "
"2*(b - c)*exp(a*t))/(t**2*(-a + b)*(-a + c)*(b - c))")
mgf = moment_generating_function(Uniform('x', a, b))(t)
assert mgf == (-exp(a * t) + exp(b * t)) / (t * (-a + b))
mgf = moment_generating_function(UniformSum('x', a))(t)
assert mgf == ((exp(t) - 1) / t)**a
mgf = moment_generating_function(WignerSemicircle('x', a))(t)
assert mgf == 2 * besseli(1, a * t) / (a * t)
# Numeric tests
mgf = moment_generating_function(Beta('x', 1, 1))(t)
assert mgf.diff(t).subs(t, 1) == hyper((2, ), (3, ), 1) / 2
mgf = moment_generating_function(Chi('x', 1))(t)
assert mgf.diff(t).subs(t, 1) == sqrt(2) * hyper(
(1, ), (Rational(3, 2), ), S.Half) / sqrt(pi) + hyper(
(Rational(3, 2), ),
(Rational(3, 2), ), S.Half) + 2 * sqrt(2) * hyper(
(2, ), (Rational(5, 2), ), S.Half) / (3 * sqrt(pi))
mgf = moment_generating_function(ChiSquared('x', 1))(t)
assert mgf.diff(t).subs(t, 1) == I
mgf = moment_generating_function(Erlang('x', 1, 1))(t)
assert mgf.diff(t).subs(t, 0) == 1
mgf = moment_generating_function(ExGaussian("x", 0, 1, 1))(t)
assert mgf.diff(t).subs(t, 2) == -exp(2)
mgf = moment_generating_function(Exponential('x', 1))(t)
assert mgf.diff(t).subs(t, 0) == 1
mgf = moment_generating_function(Gamma('x', 1, 1))(t)
assert mgf.diff(t).subs(t, 0) == 1
mgf = moment_generating_function(Gumbel('x', 1, 1))(t)
assert mgf.diff(t).subs(t, 0) == EulerGamma + 1
mgf = moment_generating_function(Gompertz('x', 1, 1))(t)
assert mgf.diff(t).subs(t, 1) == -e * meijerg(((), (1, 1)),
((0, 0, 0), ()), 1)
mgf = moment_generating_function(Laplace('x', 1, 1))(t)
assert mgf.diff(t).subs(t, 0) == 1
mgf = moment_generating_function(Logistic('x', 1, 1))(t)
assert mgf.diff(t).subs(t, 0) == beta(1, 1)
mgf = moment_generating_function(Normal('x', 0, 1))(t)
assert mgf.diff(t).subs(t, 1) == exp(S.Half)
mgf = moment_generating_function(Pareto('x', 1, 1))(t)
assert mgf.diff(t).subs(t, 0) == expint(1, 0)
mgf = moment_generating_function(QuadraticU('x', 1, 2))(t)
assert mgf.diff(t).subs(t, 1) == -12 * e - 3 * exp(2)
mgf = moment_generating_function(RaisedCosine('x', 1, 1))(t)
assert mgf.diff(t).subs(t, 1) == -2*e*pi**2*sinh(1)/\
(1 + pi**2)**2 + e*pi**2*cosh(1)/(1 + pi**2)
mgf = moment_generating_function(Rayleigh('x', 1))(t)
assert mgf.diff(t).subs(t, 0) == sqrt(2) * sqrt(pi) / 2
mgf = moment_generating_function(Triangular('x', 1, 3, 2))(t)
assert mgf.diff(t).subs(t, 1) == -e + exp(3)
mgf = moment_generating_function(Uniform('x', 0, 1))(t)
assert mgf.diff(t).subs(t, 1) == 1
mgf = moment_generating_function(UniformSum('x', 1))(t)
assert mgf.diff(t).subs(t, 1) == 1
mgf = moment_generating_function(WignerSemicircle('x', 1))(t)
assert mgf.diff(t).subs(t, 1) == -2*besseli(1, 1) + besseli(2, 1) +\
besseli(0, 1)
#
# So, the only speaker strategies that can strictly maximize expected utilities send a single message per state for all states. Since mixing signals is strictly worse, only pure strategies can be components of strict Nash equilibria and thus evolutionarily stable strategies. Thus, only speaker strategies that partition the state space in the manner described in the text can be components of evolutionarily stable strategies.
# Moving on to our analysis, first, we import `sympy`, which is a library for symbolic math in `python`.
# In[8]:
from sympy import *
from sympy.stats import Beta, density
# Next, we define the symbols that we'll use in constructing the utility functions and the prior probability.
# In[9]:
t, t_0, m_0, m_1, a_0, a_1, b = symbols('t t_0 m_0 m_1 a_0 a_1 b')
T = Beta("t", 1, 2)
# Then, we build the actual utility functions and the respective expected utilities. Note that we can save a bit of effort by noting that the speaker's utility is defined by the hearer's response and the hearer's utility is the speaker's where $b=0$.
# In[10]:
Utility_S_0 = 1 - (a_0 - t - (1 - t) * b)**2
Utility_S_1 = 1 - (a_1 - t - (1 - t) * b)**2
# From the utility functions we can calculate the expected utilities.
# In[11]:
E_Utility_S = integrate(Utility_S_0 * density(T)(t),
(t, 0, t_0)) + integrate(Utility_S_1 * density(T)(t),
(t, t_0, 1))