def abeff_trianglebound(N, y, t, cond):
sigma1 = 0.5 * (1 + y)
sum1, sum2, sum3, sum5 = [0.0 for _ in range(4)]
b1 = 1
a1 = mp.power(N, -0.4)
xN = 4 * mp.pi() * N * N - mp.pi() * t / 4.0
xNp1 = 4 * mp.pi() * (N + 1) * (N + 1) - mp.pi() * t / 4.0
delta = mp.pi() * y / (2 * (xN - 6 - (14 + 2 * y) / mp.pi())) + 2 * y * (
7 + y) * mp.log(abs(1 + y + 1j * xNp1) / (4 * mp.pi)) / (xN * xN)
expdelta = mp.exp(delta)
for n in range(1, 30 * N + 1):
nf = float(n)
denom = mp.power(nf, sigma1 + (t / 4.0) * mp.log(N * N))
common1 = mp.exp((t / 4.0) * mp.power(mp.log(nf), 2))
common2 = common1 * mp.power(nf / N, y) * expdelta * mp.exp(
t * y * mp.log(n) / (2 * (xN - 6)))
bn, bn2, bn3, bn5 = [
common1 * abs(cond[n][2 * i - 1]) for i in range(1, 5)
]
an, an2, an3, an5 = [
common2 * abs(cond[n][2 * i]) for i in range(1, 5)
]
sum1 += (bn + an) / denom
sum2 += (bn2 + an2) / denom
sum3 += (bn3 + an3) / denom
sum5 += (bn5 + an5) / denom
return [N, expdelta] + [2 - j for j in [sum1, sum2, sum3, sum5]]
def abtoyx_e3(z, t):
x = z.real
xdash = x + mp.pi() * t / 4.0
y = z.imag
sigma1, sigma2 = 0.5 * (1 + y), 0.5 * (1 - y)
s1, s2 = sigma1 + 0.5j * xdash, sigma2 + 0.5j * xdash
N = int(mp.sqrt(0.25 * x / mp.pi()))
sum1_L, sum1_R = 0.0, 0.0
factor2 = 1 - 1 / mp.power(2.0, s1 + (t / 4.0) * mp.log(N * N / 2.0))
factor3 = 1 - 1 / mp.power(3.0, s1 + (t / 4.0) * mp.log(N * N / 2.0))
factorN = mp.power(N, -0.4)
for n in range(1, N + 1):
n = float(n)
sum1_L += mp.power(n, -1 * s1 - (t / 4.0) * mp.log(N * N / n))
sum1_R += mp.power(n, -1 * s2 - (t / 4.0) * mp.log(N * N / n))
sum1_R = sum1_R * factorN
sum12_L, sum12_R = sum1_L * factor2, sum1_R * factor2
sum123_L, sum123_R = sum12_L * factor3, sum12_R * factor3
absum1_L, absum1_R, absum12_L, absum12_R, absum123_L, absum123_R = abs(
sum1_L), abs(sum1_R), abs(sum12_L), abs(sum12_R), abs(sum123_L), abs(
sum123_R)
abdiff1, abdiff12, abdiff123 = absum1_L - absum1_R, absum12_L - absum12_R, absum123_L - absum123_R
return [
sum1_L, sum1_R, absum1_L, absum1_R, abdiff1, sum12_L, sum12_R,
absum12_L, absum12_R, abdiff12, sum123_L, sum123_R, absum123_L,
absum123_R, abdiff123
]
def FreeFermions(eigvec, subsystem, FermiVector): r=range(FermiVector) Cij=mp.matrix([[mp.fsum([eigvec[i,k]*eigvec[j,k] for k in r]) for i in subsystem] for j in subsystem]) C_eigval=mp.eigsy(Cij, eigvals_only=True) EH_eigval=mp.matrix([mp.log(mp.fdiv(mp.fsub(mp.mpf(1.0),x),x)) for x in C_eigval]) S=mp.re(mp.fsum([mp.log(mp.mpf(1.0)+mp.exp(-x))+mp.fdiv(x,mp.exp(x)+mp.mpf(1.0)) for x in EH_eigval])) return(S)
def abeff_lemmabound(N, y, t, cond):
sigma1 = 0.5 * (1 + y)
sum1, sum2, sum3, sum5 = [0.0 for _ in range(4)]
b1 = 1
a1 = mp.power(N, -0.4)
xN = 4 * mp.pi() * N * N - mp.pi() * t / 4.0
xNp1 = 4 * mp.pi() * (N + 1) * (N + 1) - mp.pi() * t / 4.0
delta = mp.pi() * y / (2 * (xN - 6 - (14 + 2 * y) / mp.pi())) + 2 * y * (
7 + y) * mp.log(abs(1 + y + 1j * xNp1) / (4 * mp.pi)) / (xN * xN)
expdelta = mp.exp(delta)
for n in range(2, 30 * N + 1):
nf = float(n)
denom = mp.power(nf, sigma1 + (t / 4.0) * mp.log(N * N))
#print([cond[n][i] for i in range(1,9)])
common1 = mp.exp((t / 4.0) * mp.power(mp.log(nf), 2))
common2 = common1 * mp.power(nf / N, y)
common3 = expdelta * (mp.exp(t * y * mp.log(n) / (2 * (xN - 6))) - 1)
bn, bn2, bn3, bn5 = [common1 * cond[n][2 * i - 1] for i in range(1, 5)]
an, an2, an3, an5 = [common2 * cond[n][2 * i] for i in range(1, 5)]
en, en2, en3, en5 = an * common3, an2 * common3, an3 * common3, an5 * common3
sum1 += (en + max(
(1 - a1) * abs(bn + an) / (1 + a1), abs(bn - an))) / denom
sum2 += (en2 + max(
(1 - a1) * abs(bn2 + an2) / (1 + a1), abs(bn2 - an2))) / denom
sum3 += (en3 + max(
(1 - a1) * abs(bn3 + an3) / (1 + a1), abs(bn3 - an3))) / denom
sum5 += (en5 + max(
(1 - a1) * abs(bn5 + an5) / (1 + a1), abs(bn5 - an5))) / denom
return [N, expdelta] + [1 - a1 - j for j in [sum1, sum2, sum3, sum5]]
def cumulants(self, gamma):
"""
Compute
.. math::
\Lambda(\gamma) = \log\left(\int_{\mathbb{R}^n}
e^{\gamma \|z\|^2_2} F(dz)\right)
as well as its first two derivatives with respect to $\gamma$,
where $F$ is the empirical distribution of `self.sample`.
"""
norm_squared = self.sample
M = norm_squared.mean()
M0 = float(np.mean([mp.exp(gamma * (ns - M)) for ns in norm_squared]))
M0 *= mp.exp(gamma * M)
M1 = np.mean([
np.exp(float(gamma * ns + np.log(ns) - mp.log(M0)))
for ns in norm_squared
])
M2 = np.mean([
np.exp(float(gamma * ns + 2 * np.log(ns) - mp.log(M0)))
for ns in norm_squared
])
return M0, M1, (M2 - M1**2)
def D_transform(j0,j1,j2):
d = mp.mpf(10**(-mp.dps+1))
z0,z1,z2,z3 = 1/(j0*j1*j2),j1*j0/j2,j2*j0/j1,j1*j2/j0
c0,c1,c2,c3 = z0+z1+z2+z3,z0+z1-z2-z3,z0+z2-z1-z3,z0+z3-z1-z2
t1 = mp.sqrt(c2+0j)*mp.sqrt(c3+0j)/(mp.sqrt(c0+0j)*mp.sqrt(c1+0j))
t2 = mp.sqrt(c1+0j)*mp.sqrt(c3+0j)/(mp.sqrt(c0+0j)*mp.sqrt(c2+0j))
t3 = mp.sqrt(c1+0j)*mp.sqrt(c2+0j)/(mp.sqrt(c0+0j)*mp.sqrt(c3+0j))
if abs(1+t1)<d:
t1 = -1+d
if abs(1+t2)<d:
t2 = -1+d
if abs(1+t3)<d:
t3 = -1+d
j1p = d + mp.sqrt((1-t1)/(1+t1))
j2p = d + mp.sqrt((1-t2)/(1+t2))
jD = d + mp.sqrt((1-t3)/(1+t3))
dF = mp.log(z0/(j1p*j2p*jD+1/(j1p*j2p*jD))+0j)+mp.log(jD+1/jD)
return j1p,j2p,dF
def _solve_expx_x_logx(tau, tol, max_steps=10):
'''Solves the equation
log(pi/tau) = pi/2 * exp(x) - x - log(x)
approximately using Newton's method. The approximate solution is guaranteed
to overestimate.
'''
x = mp.log(2 / mp.pi * mp.log(mp.pi / tau))
# x = mp.log(tau/mp.pi) - mp.lambertw(-tau/2, -1))
# x = mp.mpf(1)/2 \
# - mp.log(mp.sqrt(mp.pi/tau)) \
# - mp.lambertw(-mp.sqrt(mp.exp(1)*mp.pi*tau)/4, -1)
def f0(x):
return mp.pi / 2 * mp.exp(x) - x - mp.log(x * mp.pi / tau)
def f1(x):
return mp.pi / 2 * mp.exp(x) - 1 - mp.mpf(1) / x
f0x = f0(x)
success = False
# At least one step is performed. This is required for the guarantee of
# overestimation.
for _ in range(max_steps):
x -= f0x / f1(x)
f0x = f0(x)
if abs(f0x) < tol:
success = True
break
assert success
return x
def Nt(t, T):
'''Estimates number of H_t zeroes expected to be found upto height T'''
t, T = mp.mpf(t), mp.mpf(T)
Tsmall = T / (4 * mp.pi())
N0 = Tsmall * mp.log(Tsmall) - Tsmall
extra = (t / 16.0) * mp.log(Tsmall)
return (N0 + extra).real
def kill_vertical(E, j_matrix, x, y):
y0, y1 = y, y + 1
j0 = j_matrix[y0 * L + x, y1 * L + x]
j_matrix[y0 * L + x, y1 * L + x], j_matrix[y1 * L + x, y0 * L + x] = 1, 1
dF = mp.log(j0 + 1 / j0) - mp.log(2)
E += dF
return E, j_matrix
def Ht_Effective(z, t):
"""
This uses the effective approximation of H_t from Terry's blog
:param z: point at which H_t is computed
:param t: the "time" parameter
:return: H_t as a sum of two terms that are analogous to A and B, but now also with an efffective error bound (returned as percentage of |H_t|
"""
z, t = mp.mpc(z), mp.mpc(t)
sigma = (1 - z.imag) / 2.0
T = (z.real) / 2.0
Tdash = T + t * mp.pi() / 8.0
s1 = sigma + 1j * T
s2 = 1 - sigma + 1j * T
N = int((mp.sqrt(Tdash / (2 * mp.pi()))).real)
alph1 = alpha1(s1)
alph2 = alpha1(s2).conjugate()
A0_expo = (t / 4.0) * alph1 * alph1
B0_expo = (t / 4.0) * alph2 * alph2
H01_est1 = H01(s1)
H01_est2 = H01(s2).conjugate()
#begin main estimate block
A0 = mp.exp(A0_expo) * H01_est1
B0 = mp.exp(B0_expo) * H01_est2
A_sum = 0.0
B_sum = 0.0
for n in range(1, N + 1):
A_sum += 1 / mp.power(n, s1 + (t / 2.0) * alph1 -
(t / 4.0) * mp.log(n))
B_sum += 1 / mp.power(
n, 1 - s1 + (t / 2.0) * alph2 - (t / 4.0) * mp.log(n))
A = A0 * A_sum
B = B0 * B_sum
H = (A + B) / 8.0
#end main estimate block
#begin error block
A0_err_expo = (t / 4.0) * (abs(alph1)**2) #A0_expo.real may also work
B0_err_expo = (t / 4.0) * (abs(alph2)**2) #B0_expo.real may also work
epserr_1 = mp.exp(A0_err_expo) * abs(H01_est1) * abs(eps_err(s1, t)) / (
(T - 3.33) * 8.0)
epserr_2 = mp.exp(B0_err_expo) * abs(H01_est2) * abs(eps_err(s2, t)) / (
(T - 3.33) * 8.0)
epserr = epserr_1 + epserr_2
C0 = mp.sqrt(mp.pi()) * mp.exp(-1 * (t / 64.0) * (mp.pi()**2)) * mp.power(
Tdash, 1.5) * mp.exp(-1 * mp.pi() * T / 4.0)
C = C0 * vwf_err(s1, t) / 8.0
toterr = epserr + C
#print(epserr_1, epserr_2, C0, vwf_err(s1, t), C, toterr.real)
#end error block
if z.imag == 0: return (H.real, toterr.real / abs(H.real))
else: return (H, toterr.real / abs(H))
def FreeFermions(subsystem, C):
C = mp.matrix([[C[x, y] for x in subsystem] for y in subsystem])
C_eigval = mp.eigh(C, eigvals_only=True)
EH_eigval = mp.matrix(
[mp.log(mp.fdiv(mp.fsub(mp.mpf(1.0), x), x)) for x in C_eigval])
S = mp.re(
mp.fsum([
mp.log(mp.mpf(1.0) + mp.exp(-x)) + mp.fdiv(x,
mp.exp(x) + mp.mpf(1.0))
for x in EH_eigval
]))
return (S)
def abtoybound(N, y, t, cond):
sigma1, sigma2 = 0.5 * (1 + y), 0.5 * (1 - y)
sum1_L, sum1_R, sum12_L, sum12_R, sum123_L, sum123_R, sum1235_L, sum1235_R = [
0.0 for _ in range(8)
]
ddxsum1_L, ddxsum1_R, ddxsum12_L, ddxsum12_R, ddxsum123_L, ddxsum123_R, ddxsum1235_L, ddxsum1235_R = [
0.0 for _ in range(8)
]
factorN = 1 / mp.power(N, 0.4)
for n in range(1, 30 * N + 1):
nf = float(n)
denom1 = mp.power(nf, sigma1 + (t / 4.0) * mp.log(N * N / nf))
denom2 = mp.power(nf, sigma2 + (t / 4.0) * mp.log(N * N / nf))
term1_L = abs(cond[n][1] / denom1)
term1_R = abs(cond[n][2] / denom2)
sum1_L += term1_L
sum1_R += term1_R
ddxsum1_L += mp.log(n) * term1_L
ddxsum1_R += mp.log(n) * term1_R
term12_L = abs(cond[n][3] / denom1)
term12_R = abs(cond[n][4] / denom2)
sum12_L += term12_L
sum12_R += term12_R
ddxsum12_L += mp.log(n) * term12_L
ddxsum12_R += mp.log(n) * term12_R
term123_L = abs(cond[n][5] / denom1)
term123_R = abs(cond[n][6] / denom2)
sum123_L += term123_L
sum123_R += term123_R
ddxsum123_L += mp.log(n) * term123_L
ddxsum123_R += mp.log(n) * term123_R
term1235_L = abs(cond[n][7] / denom1)
term1235_R = abs(cond[n][8] / denom2)
sum1235_L += term1235_L
sum1235_R += term1235_R
ddxsum1235_L += mp.log(n) * term1235_L
ddxsum1235_R += mp.log(n) * term1235_R
sum1_L, sum12_L, sum123_L, sum1235_L = sum1_L - 1, sum12_L - 1, sum123_L - 1, sum1235_L - 1
sum1_R, sum12_R, sum123_R, sum1235_R = sum1_R * factorN, sum12_R * factorN, sum123_R * factorN, sum1235_R * factorN
ddxsum1_L, ddxsum12_L, ddxsum123_L, ddxsum1235_L = 0.5 * ddxsum1_L, 0.5 * ddxsum12_L, 0.5 * ddxsum123_L, 0.5 * ddxsum1235_L
ddxsum1_R, ddxsum12_R, ddxsum123_R, ddxsum1235_R = 0.5 * ddxsum1_R * factorN, 0.5 * ddxsum12_R * factorN, 0.5 * ddxsum123_R * factorN, 0.5 * ddxsum1235_R * factorN
abdiff1, ddxsum1 = 1 - sum1_L - sum1_R, ddxsum1_L + ddxsum1_R
abdiff12, ddxsum12 = 1 - sum12_L - sum12_R, ddxsum12_L + ddxsum12_R
abdiff123, ddxsum123 = 1 - sum123_L - sum123_R, ddxsum123_L + ddxsum123_R
abdiff1235, ddxsum1235 = 1 - sum1235_L - sum1235_R, ddxsum1235_L + ddxsum1235_R
return [
sum1_L, sum1_R, abdiff1, ddxsum1_L, ddxsum1_R, ddxsum1, sum12_L,
sum12_R, abdiff12, ddxsum12_L, ddxsum12_R, ddxsum12, sum123_L,
sum123_R, abdiff123, ddxsum123_L, ddxsum123_R, ddxsum123, sum1235_L,
sum1235_R, abdiff1235, ddxsum1235_L, ddxsum1235_R, ddxsum1235
]
def FreeFermions(subsystem,
C_t): #implements free fermion technique by peschel
C = mp.matrix([[C_t[x, y] for x in subsystem] for y in subsystem])
C_eigval = mp.eigh(C, eigvals_only=True)
EH_eigval = mp.matrix(
[mp.log(mp.fdiv(mp.fsub(mp.mpf(1.0), x), x)) for x in C_eigval])
S = mp.re(
mp.fsum([
mp.log(mp.mpf(1.0) + mp.exp(-x)) + mp.fdiv(x,
mp.exp(x) + mp.mpf(1.0))
for x in EH_eigval
]))
return (S)
def calc_lmsr_marginal_price(token_count, token_index, net_outcome_tokens_sold,
funding):
mp.dps = 100
mp.pretty = True
b = mpf(funding) / mp.log(len(net_outcome_tokens_sold))
result = b * mp.log(
sum(
mp.exp(share_count / b + token_count / b)
for share_count in net_outcome_tokens_sold) - sum(
mp.exp(share_count / b)
for index, share_count in enumerate(net_outcome_tokens_sold) if
index != token_index)) - net_outcome_tokens_sold[token_index]
return result
def ln_I(k, n, t, rho, X, W):
N = len(X)
y0 = find_y0(k, n, t, rho)
fy0 = f(y0, k, n, t, rho)
fppy0 = d2fdy2(y0, k, n, t, rho)
#X,W = scipy.special.he_roots(N)
r = lambda y: f(y, k, n, t, rho) - fy0 + ((y - y0)**2) * (abs(fppy0) / 2.)
ln_ans = mp.log(
sum([
W[i] * mp.exp(r(mp.sqrt(1. / abs(fppy0)) * X[i] + y0))
for i in range(N)
]))
ln_ans = ln_ans + fy0 + 0.5 * mp.log(1 - rho) - 0.5 * mp.log(
2 * mp.pi * rho * abs(fppy0))
return (ln_ans)
def kill_horizontal(E, j_matrix, x, y):
if x > 0:
j1 = j_matrix[y * L + x, y * L + x + 1]
j2 = j_matrix[y * L + x - 1, y * L + x]
jp, dF = L_transform(j1, j2)
j_matrix[y * L + x + 1, y * L + x], j_matrix[y * L + x,
y * L + x + 1] = 1, 1
j_matrix[y * L + x - 1, y * L + x], j_matrix[y * L + x,
y * L + x - 1] = jp, jp
E += dF - mp.log(2)
if x == 0:
j0 = j_matrix[0, 1]
j_matrix[0, 1], j_matrix[1, 0] = 1, 1
E += mp.log(j0 + 1 / j0) - mp.log(2)
return E, j_matrix
def z_Zolotarev(N, x, m):
r"""
Function to evaluate the Zolotarev polynomial (eq 1, [McNamara93]_).
:param N: Order of the Zolotarev polynomial
:param x: The argument at which one would like to evaluate the Zolotarev polynomial
:param m: m is the elliptic parameter (not the modulus k and not the nome q)
:rtype: Returns a float, the value of Zolotarev polynomial at x
"""
M = -ellipk(m) / N
x3 = ellipfun('sn', u=-M, m=m)
xbar = x3 * mp.sqrt(
(x**2 - 1) / (x**2 - x3**2)) # rearranged eq 21, [Levy70]_
u = ellipf(mp.asin(xbar),
m) # rearranged eq 20, [Levy70]_, asn(x) = F(asin(x)|m)
f = mp.cosh((N / 2) * mp.log(z_eta(M + u, m) / z_eta(M - u, m)))
if f.imag / f.real > 1e-10:
print("imaginary part of the Zolotarev function is not negligible!")
print("f_imaginary = ", f.imag)
else:
if (x > 0): # no idea why I am doing this ... anyhow, it seems working
f = -f.real
else:
f = f.real
return f
def test_svd_test_case():
# a test case from Golub and Reinsch
# (see wilkinson/reinsch: handbook for auto. comp., vol ii-linear algebra, 134-151(1971).)
eps = mp.exp(0.8 * mp.log(mp.eps))
a = [[22, 10, 2, 3, 7],
[14, 7, 10, 0, 8],
[-1, 13, -1, -11, 3],
[-3, -2, 13, -2, 4],
[ 9, 8, 1, -2, 4],
[ 9, 1, -7, 5, -1],
[ 2, -6, 6, 5, 1],
[ 4, 5, 0, -2, 2]]
a = mp.matrix(a)
b = mp.matrix([mp.sqrt(1248), 20, mp.sqrt(384), 0, 0])
S = mp.svd_r(a, compute_uv = False)
S -= b
assert mp.mnorm(S) < eps
S = mp.svd_c(a, compute_uv = False)
S -= b
assert mp.mnorm(S) < eps
def run_eig(A, verbose = 0):
if verbose > 1:
print("original matrix (eig):\n", A)
n = A.rows
E, EL, ER = mp.eig(A, left = True, right = True)
if verbose > 1:
print("E:\n", E)
print("EL:\n", EL)
print("ER:\n", ER)
eps = mp.exp(0.8 * mp.log(mp.eps))
err0 = 0
for i in xrange(n):
B = A * ER[:,i] - E[i] * ER[:,i]
err0 = max(err0, mp.mnorm(B))
B = EL[i,:] * A - EL[i,:] * E[i]
err0 = max(err0, mp.mnorm(B))
err0 /= n * n
if verbose > 0:
print("difference (E):", err0)
assert err0 < eps
def floor_power(x, n=2):
"""
Return the value sign(x) * n^k such that n^k is the largest value <= |x|
for integer k.
"""
x = abs(x)
return mp.power(n, mp.floor(mp.log(x, n)))
def next_power(x, n=2):
"""
Return the value sign(x) * n^k such that n^k is the smallest value > |x|
for integer k.
"""
x = abs(x)
return mp.power(n, mp.floor(mp.log(x, n)) + 1)
def entropy_pf(Lph,Tph,meas,G,prec=15,q=2,dps=200):
# setting digit precision
mp.dps = dps
# =============================================================================
# Evaluation of the entropy using partition function
# Lph,Tph -- physical size (number of qubits) and time (circuit depth)
# G -- (large) parameter controling the gap between relevant and irrelevant states
# prec -- number of digits in the answer
# q -- qudit dimension
# =============================================================================
# -- define effective inverse temperature
beta = mp.log((q**2+1)/q)
# -- dimension of the effective lattice
L,T = int(Lph/2)+(Lph+1)%2,Tph+1
# -- corresponding couplings for the numerator and denominator
J_matrix1,J_matrix2 = generate_lattice_couplings(L,T,meas,G,beta)
# -- including the temperature
J_matrix1 = -beta*mp.matrix(J_matrix1.tolist())
J_matrix2 = -beta*mp.matrix(J_matrix2.tolist())
# -- evaluation of the entropy
entropy = -log_partition_function(J_matrix1,L,T).real\
+log_partition_function(J_matrix2,L,T).real
return mp.nstr(entropy,prec)
def run_eigsy(A, verbose = False):
if verbose:
print("original matrix:\n", str(A))
D, Q = mp.eigsy(A)
B = Q * mp.diag(D) * Q.transpose()
C = A - B
E = Q * Q.transpose() - mp.eye(A.rows)
if verbose:
print("eigenvalues:\n", D)
print("eigenvectors:\n", Q)
NC = mp.mnorm(C)
NE = mp.mnorm(E)
if verbose:
print("difference:", NC, "\n", C, "\n")
print("difference:", NE, "\n", E, "\n")
eps = mp.exp( 0.8 * mp.log(mp.eps))
assert NC < eps
assert NE < eps
return NC
def log_likelihood(self, x, S=10):
# define the posterior q(z|x) / encode x into q(z|x)
qz = self.posterior(x)
# define the prior p(z)
pz = self.prior(batch_size=x.size(0))
# sample S samples from the posterior per data point x
z = qz.rsample([S]) # [S, batchsize, latentdim]
# define the observation model p(x|z) = B(x | g(z))
px = self.observation_model(z)
# Calculating Monte Carlo Estimate of log likelihood
sum_log_lik = px.log_prob(x).sum(-1) + pz.log_prob(z).sum(
-1) - qz.log_prob(z).sum(-1)
log_lik = torch.zeros(x.shape[0])
for i in range(x.shape[0]):
tmp = mp.log(
sum([mp.exp(t)
for t in sum_log_lik[:, i].detach().numpy()]) / S)
log_lik[i] = float(tmp)
ave_log_lik = log_lik.mean()
n_in_ave = x.shape[0]
return {
"log_like": log_lik,
"average_log_like": ave_log_lik,
"n": n_in_ave
}
def run_hessenberg(A, verbose=0):
if verbose > 1:
print("original matrix (hessenberg):\n", A)
n = A.rows
Q, H = mp.hessenberg(A)
if verbose > 1:
print("Q:\n", Q)
print("H:\n", H)
B = Q * H * Q.transpose_conj()
eps = mp.exp(0.8 * mp.log(mp.eps))
err0 = 0
for x in xrange(n):
for y in xrange(n):
err0 += abs(A[y, x] - B[y, x])
err0 /= n * n
err1 = 0
for x in xrange(n):
for y in xrange(x + 2, n):
err1 += abs(H[y, x])
if verbose > 0:
print("difference (H):", err0, err1)
if verbose > 1:
print("B:\n", B)
assert err0 < eps
assert err1 == 0
def test_levin_2():
# [2] A. Sidi - "Pratical Extrapolation Methods" p.373
mp.dps = 17
z = mp.mpf(10)
eps = mp.mpf(mp.eps)
with mp.extraprec(2 * mp.prec):
L = mp.levin(method="sidi", variant="t")
n = 0
while 1:
s = (-1)**n * mp.fac(n) * z**(-n)
v, e = L.step(s)
n += 1
if e < eps:
break
if n > 1000: raise RuntimeError("iteration limit exceeded")
eps = mp.exp(0.9 * mp.log(eps))
exact = mp.quad(lambda x: mp.exp(-x) / (1 + x / z), [0, mp.inf])
# there is also a symbolic expression for the integral:
# exact = z * mp.exp(z) * mp.expint(1,z)
err = abs(v - exact)
assert err < eps
w = mp.nsum(lambda n: (-1)**n * mp.fac(n) * z**(-n), [0, mp.inf],
method="sidi",
levin_variant="t")
assert err < eps
def Ht_AFE_B(z, t):
"""
This is the much more accurate approx functional eqn posted by Terry at
https://terrytao.wordpress.com/2018/02/02/polymath15-second-thread-generalising-the-riemann-siegel-approximate-functional-equation/#comment-492182
:param z: point at which H_t is computed
:param t: the "time" parameter
:return: the B part in Ht
"""
z, t = mp.mpc(z), mp.mpc(t)
s = (1 + 1j * z.real - z.imag) / 2
tau = mp.sqrt(s.imag / (2 * mp.pi()))
M = int(tau)
B_pre = (1 / 16.0) * s * (s - 1) * mp.power(mp.pi(), 0.5 *
(s - 1)) * mp.gamma(0.5 *
(1 - s))
B_sum = 0.0
for m in range(1, M + 1):
if t.real > 0:
B_sum += mp.exp(
(t / 16.0) * mp.power(mp.log(
(5 - s) / (2 * mp.pi() * m * m)), 2)) / mp.power(m, 1 - s)
else:
B_sum += 1 / mp.power(m, 1 - s)
return B_pre * B_sum
def w(sigma, t, T0dash):
wterm1 = 1 + (sigma**2) / (T0dash**2)
wterm2 = 1 + ((1 - sigma)**2) / (T0dash**2)
wterm3 = (sigma - 1) * mp.log(wterm1) / 4.0 + nonnegative(
(T0dash / 2.0) * mp.atan(sigma / T0dash) -
sigma / 2.0) + 1 / (12.0 * (T0dash - 0.33))
return mp.sqrt(wterm1) * mp.sqrt(wterm2) * mp.exp(wterm3)
def run_hessenberg(A, verbose = 0):
if verbose > 1:
print("original matrix (hessenberg):\n", A)
n = A.rows
Q, H = mp.hessenberg(A)
if verbose > 1:
print("Q:\n",Q)
print("H:\n",H)
B = Q * H * Q.transpose_conj()
eps = mp.exp(0.8 * mp.log(mp.eps))
err0 = 0
for x in xrange(n):
for y in xrange(n):
err0 += abs(A[y,x] - B[y,x])
err0 /= n * n
err1 = 0
for x in xrange(n):
for y in xrange(x + 2, n):
err1 += abs(H[y,x])
if verbose > 0:
print("difference (H):", err0, err1)
if verbose > 1:
print("B:\n", B)
assert err0 < eps
assert err1 == 0
def Ht_AFE_A(z, t):
"""
This is the much more accurate approx functional eqn posted by Terry at
https://terrytao.wordpress.com/2018/02/02/polymath15-second-thread-generalising-the-riemann-siegel-approximate-functional-equation/#comment-492182
:param z: point at which H_t is computed
:param t: the "time" parameter
:return: the A part in Ht
"""
z, t = mp.mpc(z), mp.mpc(t)
s = (1 + 1j * z.real - z.imag) / 2
tau = mp.sqrt(s.imag / (2 * mp.pi()))
N = int(tau)
A_pre = (1/16) * s * (s-1) \
* mp.power(mp.pi(), -1*s/2) * mp.gamma(s/2)
A_sum = 0.0
for n in range(1, N + 1):
if t.real > 0:
A_sum += mp.exp(
(t / 16) * mp.power(mp.log(
(s + 4) / (2 * mp.pi() * n * n)), 2)) / mp.power(n, s)
else:
A_sum += 1 / mp.power(n, s)
return A_pre * A_sum
def test_levin_3():
mp.dps = 17
z = mp.mpf(2)
eps = mp.mpf(mp.eps)
with mp.extraprec(
7 * mp.prec
): # we need copious amount of precision to sum this highly divergent series
L = mp.levin(method="levin", variant="t")
n, s = 0, 0
while 1:
s += (-z)**n * mp.fac(4 * n) / (mp.fac(n) * mp.fac(2 * n) * (4**n))
n += 1
v, e = L.step_psum(s)
if e < eps:
break
if n > 1000: raise RuntimeError("iteration limit exceeded")
eps = mp.exp(0.8 * mp.log(eps))
exact = mp.quad(lambda x: mp.exp(-x * x / 2 - z * x**4),
[0, mp.inf]) * 2 / mp.sqrt(2 * mp.pi)
# there is also a symbolic expression for the integral:
# exact = mp.exp(mp.one / (32 * z)) * mp.besselk(mp.one / 4, mp.one / (32 * z)) / (4 * mp.sqrt(z * mp.pi))
err = abs(v - exact)
assert err < eps
w = mp.nsum(lambda n: (-z)**n * mp.fac(4 * n) /
(mp.fac(n) * mp.fac(2 * n) * (4**n)), [0, mp.inf],
method="levin",
levin_variant="t",
workprec=8 * mp.prec,
steps=[2] + [1 for x in xrange(1000)])
err = abs(v - w)
assert err < eps
def run_eigsy(A, verbose=False):
if verbose:
print("original matrix:\n", str(A))
D, Q = mp.eigsy(A)
B = Q * mp.diag(D) * Q.transpose()
C = A - B
E = Q * Q.transpose() - mp.eye(A.rows)
if verbose:
print("eigenvalues:\n", D)
print("eigenvectors:\n", Q)
NC = mp.mnorm(C)
NE = mp.mnorm(E)
if verbose:
print("difference:", NC, "\n", C, "\n")
print("difference:", NE, "\n", E, "\n")
eps = mp.exp(0.8 * mp.log(mp.eps))
assert NC < eps
assert NE < eps
return NC
def run_eig(A, verbose=0):
if verbose > 1:
print("original matrix (eig):\n", A)
n = A.rows
E, EL, ER = mp.eig(A, left=True, right=True)
if verbose > 1:
print("E:\n", E)
print("EL:\n", EL)
print("ER:\n", ER)
eps = mp.exp(0.8 * mp.log(mp.eps))
err0 = 0
for i in xrange(n):
B = A * ER[:, i] - E[i] * ER[:, i]
err0 = max(err0, mp.mnorm(B))
B = EL[i, :] * A - EL[i, :] * E[i]
err0 = max(err0, mp.mnorm(B))
err0 /= n * n
if verbose > 0:
print("difference (E):", err0)
assert err0 < eps
def test_levin_nsum():
mp.dps = 17
with mp.extraprec(mp.prec):
z = mp.mpf(10) ** (-10)
a = mp.nsum(lambda n: n**(-(1+z)), [1, mp.inf], method = "l") - 1 / z
assert abs(a - mp.euler) < 1e-10
eps = mp.exp(0.8 * mp.log(mp.eps))
a = mp.nsum(lambda n: (-1)**(n-1) / n, [1, mp.inf], method = "sidi")
assert abs(a - mp.log(2)) < eps
z = 2 + 1j
f = lambda n: mp.rf(2 / mp.mpf(3), n) * mp.rf(4 / mp.mpf(3), n) * z**n / (mp.rf(1 / mp.mpf(3), n) * mp.fac(n))
v = mp.nsum(f, [0, mp.inf], method = "levin", steps = [10 for x in xrange(1000)])
exact = mp.hyp2f1(2 / mp.mpf(3), 4 / mp.mpf(3), 1 / mp.mpf(3), z)
assert abs(exact - v) < eps
def cumulants(self, gamma):
"""
Compute
.. math::
\Lambda(\gamma) = \log\left(\int_{\mathbb{R}^n}
e^{\gamma \|z\|^2_2} F(dz)\right)
as well as its first two derivatives with respect to $\gamma$,
where $F$ is the empirical distribution of `self.sample`.
"""
norm_squared = self.sample
M = norm_squared.mean()
M0 = float(np.mean([mp.exp(gamma*(ns-M)) for ns in norm_squared]))
M0 *= mp.exp(gamma*M)
M1 = np.mean([np.exp(float(gamma*ns+np.log(ns)-mp.log(M0))) for ns in norm_squared])
M2 = np.mean([np.exp(float(gamma*ns+2*np.log(ns)-mp.log(M0))) for ns in norm_squared])
return M0, M1, (M2-M1**2)
def test_cohen_alt_0():
mp.dps = 17
AC = mp.cohen_alt()
S, s, n = [], 0, 1
while 1:
s += -((-1) ** n) * mp.one / (n * n)
n += 1
S.append(s)
v, e = AC.update_psum(S)
if e < mp.eps:
break
if n > 1000: raise RuntimeError("iteration limit exceeded")
eps = mp.exp(0.9 * mp.log(mp.eps))
err = abs(v - mp.pi ** 2 / 12)
assert err < eps
def z_Zolotarev(N, x, m):
"""Function to evaluate the Zolotarev polynomial (eq 1, [McNamara93]_)."""
M = -ellipk(m) / N
x3 = ellipfun('sn', u= -M, m=m)
xbar = x3 * mp.sqrt((x ** 2 - 1) / (x ** 2 - x3 ** 2)) # rearranged eq 21, [Levy70]_
u = ellipf(mp.asin(xbar), m) # rearranged eq 20, [Levy70]_, asn(x) = F(asin(x)|m)
f = mp.cosh((N / 2) * mp.log(z_eta(M + u, m) / z_eta(M - u, m)))
if (f.imag / f.real > 1e-10):
print "imaginary part of the Zolotarev function is not negligible!"
print "f_imaginary = ", f.imag
else:
if (x > 0): # no idea why I am doing this ... anyhow, it seems working
f = -f.real
else:
f = f.real
return f
def entropy(NMZ, NM, NZ, NZW, M, K, J, alpha, phi):
'''
compute perplexity as a function of entropy of the model
'''
AK = K * alpha
N = 0
ent = 0
for m, d in enumerate(DTM):
#print "m:", m
#print "d", d
theta = NMZ[m, :] / (M + AK)
#print theta
ent -= mp.log(np.inner(dryrun[:,m],theta))
#print "ent:", ent
N += M
return mp.exp(ent/N)
def run_schur(A, verbose = 0):
if verbose > 1:
print("original matrix (schur):\n", A)
n = A.rows
Q, R = mp.schur(A)
if verbose > 1:
print("Q:\n", Q)
print("R:\n", R)
B = Q * R * Q.transpose_conj()
C = Q * Q.transpose_conj()
eps = mp.exp(0.8 * mp.log(mp.eps))
err0 = 0
for x in xrange(n):
for y in xrange(n):
err0 += abs(A[y,x] - B[y,x])
err0 /= n * n
err1 = 0
for x in xrange(n):
for y in xrange(n):
if x == y:
C[y,x] -= 1
err1 += abs(C[y,x])
err1 /= n * n
err2 = 0
for x in xrange(n):
for y in xrange(x + 1, n):
err2 += abs(R[y,x])
if verbose > 0:
print("difference (S):", err0, err1, err2)
if verbose > 1:
print("B:\n", B)
assert err0 < eps
assert err1 < eps
assert err2 == 0
def test_levin_0():
mp.dps = 17
eps = mp.mpf(mp.eps)
with mp.extraprec(2 * mp.prec):
L = mp.levin(method = "levin", variant = "u")
S, s, n = [], 0, 1
while 1:
s += mp.one / (n * n)
n += 1
S.append(s)
v, e = L.update_psum(S)
if e < eps:
break
if n > 1000: raise RuntimeError("iteration limit exceeded")
eps = mp.exp(0.9 * mp.log(eps))
err = abs(v - mp.pi ** 2 / 6)
assert err < eps
w = mp.nsum(lambda n: 1/(n * n), [1, mp.inf], method = "levin", levin_variant = "u")
err = abs(v - w)
assert err < eps
def test_levin_1():
mp.dps = 17
eps = mp.mpf(mp.eps)
with mp.extraprec(2 * mp.prec):
L = mp.levin(method = "levin", variant = "v")
A, n = [], 1
while 1:
s = mp.mpf(n) ** (2 + 3j)
n += 1
A.append(s)
v, e = L.update(A)
if e < eps:
break
if n > 1000: raise RuntimeError("iteration limit exceeded")
eps = mp.exp(0.9 * mp.log(eps))
err = abs(v - mp.zeta(-2-3j))
assert err < eps
w = mp.nsum(lambda n: n ** (2 + 3j), [1, mp.inf], method = "levin", levin_variant = "v")
err = abs(v - w)
assert err < eps
def run_svd_r(A, full_matrices = False, verbose = True):
m, n = A.rows, A.cols
eps = mp.exp(0.8 * mp.log(mp.eps))
if verbose:
print("original matrix:\n", str(A))
print("full", full_matrices)
U, S0, V = mp.svd_r(A, full_matrices = full_matrices)
S = mp.zeros(U.cols, V.rows)
for j in xrange(min(m, n)):
S[j,j] = S0[j]
if verbose:
print("U:\n", str(U))
print("S:\n", str(S0))
print("V:\n", str(V))
C = U * S * V - A
err = mp.mnorm(C)
if verbose:
print("C\n", str(C), "\n", err)
assert err < eps
D = V * V.transpose() - mp.eye(V.rows)
err = mp.mnorm(D)
if verbose:
print("D:\n", str(D), "\n", err)
assert err < eps
E = U.transpose() * U - mp.eye(U.cols)
err = mp.mnorm(E)
if verbose:
print("E:\n", str(E), "\n", err)
assert err < eps
def test_levin_3():
mp.dps = 17
z=mp.mpf(2)
eps = mp.mpf(mp.eps)
with mp.extraprec(7*mp.prec): # we need copious amount of precision to sum this highly divergent series
L = mp.levin(method = "levin", variant = "t")
n, s = 0, 0
while 1:
s += (-z)**n * mp.fac(4 * n) / (mp.fac(n) * mp.fac(2 * n) * (4 ** n))
n += 1
v, e = L.step_psum(s)
if e < eps:
break
if n > 1000: raise RuntimeError("iteration limit exceeded")
eps = mp.exp(0.8 * mp.log(eps))
exact = mp.quad(lambda x: mp.exp( -x * x / 2 - z * x ** 4), [0,mp.inf]) * 2 / mp.sqrt(2 * mp.pi)
# there is also a symbolic expression for the integral:
# exact = mp.exp(mp.one / (32 * z)) * mp.besselk(mp.one / 4, mp.one / (32 * z)) / (4 * mp.sqrt(z * mp.pi))
err = abs(v - exact)
assert err < eps
w = mp.nsum(lambda n: (-z)**n * mp.fac(4 * n) / (mp.fac(n) * mp.fac(2 * n) * (4 ** n)), [0, mp.inf], method = "levin", levin_variant = "t", workprec = 8*mp.prec, steps = [2] + [1 for x in xrange(1000)])
err = abs(v - w)
assert err < eps
def test_levin_2():
# [2] A. Sidi - "Pratical Extrapolation Methods" p.373
mp.dps = 17
z=mp.mpf(10)
eps = mp.mpf(mp.eps)
with mp.extraprec(2 * mp.prec):
L = mp.levin(method = "sidi", variant = "t")
n = 0
while 1:
s = (-1)**n * mp.fac(n) * z ** (-n)
v, e = L.step(s)
n += 1
if e < eps:
break
if n > 1000: raise RuntimeError("iteration limit exceeded")
eps = mp.exp(0.9 * mp.log(eps))
exact = mp.quad(lambda x: mp.exp(-x)/(1+x/z),[0,mp.inf])
# there is also a symbolic expression for the integral:
# exact = z * mp.exp(z) * mp.expint(1,z)
err = abs(v - exact)
assert err < eps
w = mp.nsum(lambda n: (-1) ** n * mp.fac(n) * z ** (-n), [0, mp.inf], method = "sidi", levin_variant = "t")
assert err < eps
def z_Zolotarev(N, x, m):
r"""
Function to evaluate the Zolotarev polynomial (eq 1, [McNamara93]_).
:param N: Order of the Zolotarev polynomial
:param x: The argument at which one would like to evaluate the Zolotarev polynomial
:param m: m is the elliptic parameter (not the modulus k and not the nome q)
:rtype: Returns a float, the value of Zolotarev polynomial at x
"""
M = -ellipk(m) / N
x3 = ellipfun('sn', u= -M, m=m)
xbar = x3 * mp.sqrt((x ** 2 - 1) / (x ** 2 - x3 ** 2)) # rearranged eq 21, [Levy70]_
u = ellipf(mp.asin(xbar), m) # rearranged eq 20, [Levy70]_, asn(x) = F(asin(x)|m)
f = mp.cosh((N / 2) * mp.log(z_eta(M + u, m) / z_eta(M - u, m)))
if (f.imag / f.real > 1e-10):
print "imaginary part of the Zolotarev function is not negligible!"
print "f_imaginary = ", f.imag
else:
if (x > 0): # no idea why I am doing this ... anyhow, it seems working
f = -f.real
else:
f = f.real
return f
############ # RG58 or Jans numbers ############ a = 0.4675*10**-3; # mm b = 1.475*10**-3; # mm c = 1.8 *10**-3; # mm sigma = 5.8*10**7; # 1/Ohm*m epsPE = 1.9; f = np.logspace(5, 8, 100); omega = 2*math.pi*f; Lp = mu0/(2*math.pi)*mp.log(b/a); # ln(ra/ri) Cp = 2*math.pi*eps0*epsPE/(mp.log(b/a)); # ln(ra/ri) etac = np.sqrt(1j*omega*mu0/sigma); gammac = np.sqrt(1j*omega*mu0*sigma); # shortcuts i0 = lambda x: mp.besseli(0,x) i1 = lambda x: mp.besseli(1,x) k0 = lambda x: mp.besselk(0,x) k1 = lambda x: mp.besselk(1,x) # Zap = etac/(2*math.pi*a)*(sp.iv(0,gammac*a) / sp.iv(1,gammac*a)); ZapA = etac/(2*math.pi*a)
# NOTE: This file implements an slightly different version of li_criter. # It finds the taylor expansion coefficients of log(xi(z/(z-1)), instead if its derivative, # which is in the original li_criterion # # the following code is from # http://fredrikj.net/blog/2013/03/testing-lis-criterion/ # It uses mpmath to calculate taylor expansion of xi function # # It will produce the 1st 21 coefficients for Li-criter # [-0.69315, 0.023096, 0.046173, 0.069213, 0.092198, 0.11511, 0.13793, 0.16064, 0.18322, 0.20566, # 0.22793, 0.25003, 0.27194, 0.29363, 0.31511, 0.33634, 0.35732, 0.37803, 0.39847, 0.41862, 0.43846] # # More information about mpmath can be found at: mpmath.org # http://mpmath.org/ from mpmath import mp mp.dps = 5 mp.pretty = True xi = lambda s: (s - 1) * mp.pi ** (-0.5 * s) * mp.gamma(1 + 0.5 * s) * mp.zeta(s) # calculate 1st 21 coefficients of taylor expansion of log(xi(z/(z-1)) tmp = mp.taylor(lambda z: mp.log(xi(z / (z - 1))), 0, 20) print tmp
