In [3]:
higham(2, 30)


Out[3]:
3.9999999671099733

In [5]:
higham(2, 5)


Out[5]:
3.999999999999986

In [17]:
higham(2, 60)


Out[17]:
1.0

In [14]:
sqrt_eps = 1.5e-8
t = 2
L =

In [15]:
higham(t + sqrt_eps, L) - higham(t, L)


Out[15]:
5.999996766092863e-08

In [ ]:
def delta(f, h, t0):
    return fabs(f(t0 - h) - 2*f(t0) + f(t0 + h))

def mu(delta):
    tau1 = 100
    tau2 = .1
    ha = epsf**.25

In [2]:
import numpy as np

In [3]:
fval = np.array([0, 1, 2, 3, 4])

In [5]:
fval.size


Out[5]:
5

In [6]:
fval[0:2]


Out[6]:
array([0, 1])

In [7]:
from math import sqrt, fabs
import numpy as np

def higham(t, L):
    f = t
    for k in range(1, L + 1):
        f = sqrt(f)
    for k in range(1, L + 1):
        f = f**2
    f = f**2
    return f

#
#  Determines the noise of a function from the function values
#
#     [fnoise,level,inform] = ECnoise(nf,fval)
#
#  The user must provide the function value at nf equally-spaced points.
#  For example, if nf = 7, the user could provide
#
#     f(x-3h), f(x-2h), f(x-h), f(x), f(x+h), f(x+2h), f(x+3h)
#
#  in the array fval. Although nf >= 4 is allowed, the use of at least
#  nf = 7 function evaluations is recommended.
#
#  Noise will not be detected by this code if the function values differ
#  in the first digit.
#
#  If noise is not detected, the user should increase or decrease the
#  spacing h according to the ouput value of inform.  In most cases,
#  the subroutine detects noise with the initial value of h.
#
#  On exit:
#    fnoise is set to an estimate of the function noise;
#       fnoise is set to zero if noise is not detected.
#
#    level is set to estimates for the noise. The k-th entry is an
#      estimate from the k-th difference.
#
#    inform is set as follows:
#      inform = 1  Noise has been detected.
#      inform = 2  Noise has not been detected; h is too small.
#                  Try 100*h for the next value of h.
#      inform = 3  Noise has not been detected; h is too large.
#                  Try h/100 for the next value of h.
#
#     Argonne National Laboratory
#     Jorge More' and Stefan Wild. November 2009.

def ECNoise(nf, fval):
    level = np.zeros((nf-1))
    dsgn  = np.zeros((nf-1))
    fnoise = 0.0
    gamma = 1.0 # = gamma(0)

    # Compute the range of function values.
    fmin = np.amin(fval)
    fmax = np.amax(fval)
    if (fmax-fmin)/max(fabs(fmax), fabs(fmin)) > .1:
        inform = 3
        return fnoise, level, inform

    # Construct the difference table.
    for j in range(nf-1):
        for i in range(nf-(j+1)):
            fval[i] = fval[i+1] - fval[i]

        # h is too small only when half the function values are equal.
        if (j==0 and sum([fval[k] == 0 for k in range(nf - 1)]) >= nf/2):
            inform = 2
            return fnoise, level, inform

        gamma = 0.5*((j+1.)/(2.*(j+1.)-1.))*gamma

        # Compute the estimates for the noise level.
        level[j] = sqrt(gamma*np.mean(np.square(fval[0:nf-(j+1)])))

        # Determine differences in sign.
        emin = np.amin(fval[0:nf-(j+1)])
        emax = np.amax(fval[0:nf-(j+1)])
        if (emin*emax < 0.0):
            dsgn[j] = 1

    # Determine the noise level.
    for k in range(nf-3):
        emin = np.amin(level[k:k+3])
        emax = np.amax(level[k:k+3])
        if (emax<=4*emin and dsgn[k]):
            fnoise = level[k]
            inform = 1
            return fnoise, level, inform

    # If noise not detected then h is too large.
    inform = 3
    return fnoise, level, inform

In [8]:
fval = np.zeros((7))
sqrt_eps = 1.5e-8
initial_h = 100*sqrt_eps

for i,j in zip(range(-3,4), range(0,7)):
    fval[j] = higham(2 + i*initial_h, 30)
    

fnoise, level, inform = ECNoise(7, fval)

In [9]:
inform


Out[9]:
1

In [10]:
fnoise


Out[10]:
4.9251375751751667e-07

In [27]:
level


Out[27]:
array([  4.30033092e-06,   0.00000000e+00,   0.00000000e+00,
         0.00000000e+00,   0.00000000e+00,   0.00000000e+00])

In [ ]: