Related
I am looking for a function in Numpy or Scipy (or any rigorous Python library) that will give me the cumulative normal distribution function in Python.
Here's an example:
>>> from scipy.stats import norm
>>> norm.cdf(1.96)
0.9750021048517795
>>> norm.cdf(-1.96)
0.024997895148220435
In other words, approximately 95% of the standard normal interval lies within two standard deviations, centered on a standard mean of zero.
If you need the inverse CDF:
>>> norm.ppf(norm.cdf(1.96))
array(1.9599999999999991)
It may be too late to answer the question but since Google still leads people here, I decide to write my solution here.
That is, since Python 2.7, the math library has integrated the error function math.erf(x)
The erf() function can be used to compute traditional statistical functions such as the cumulative standard normal distribution:
from math import *
def phi(x):
#'Cumulative distribution function for the standard normal distribution'
return (1.0 + erf(x / sqrt(2.0))) / 2.0
Ref:
https://docs.python.org/2/library/math.html
https://docs.python.org/3/library/math.html
How are the Error Function and Standard Normal distribution function related?
Starting Python 3.8, the standard library provides the NormalDist object as part of the statistics module.
It can be used to get the cumulative distribution function (cdf - probability that a random sample X will be less than or equal to x) for a given mean (mu) and standard deviation (sigma):
from statistics import NormalDist
NormalDist(mu=0, sigma=1).cdf(1.96)
# 0.9750021048517796
Which can be simplified for the standard normal distribution (mu = 0 and sigma = 1):
NormalDist().cdf(1.96)
# 0.9750021048517796
NormalDist().cdf(-1.96)
# 0.024997895148220428
Adapted from here http://mail.python.org/pipermail/python-list/2000-June/039873.html
from math import *
def erfcc(x):
"""Complementary error function."""
z = abs(x)
t = 1. / (1. + 0.5*z)
r = t * exp(-z*z-1.26551223+t*(1.00002368+t*(.37409196+
t*(.09678418+t*(-.18628806+t*(.27886807+
t*(-1.13520398+t*(1.48851587+t*(-.82215223+
t*.17087277)))))))))
if (x >= 0.):
return r
else:
return 2. - r
def ncdf(x):
return 1. - 0.5*erfcc(x/(2**0.5))
To build upon Unknown's example, the Python equivalent of the function normdist() implemented in a lot of libraries would be:
def normcdf(x, mu, sigma):
t = x-mu;
y = 0.5*erfcc(-t/(sigma*sqrt(2.0)));
if y>1.0:
y = 1.0;
return y
def normpdf(x, mu, sigma):
u = (x-mu)/abs(sigma)
y = (1/(sqrt(2*pi)*abs(sigma)))*exp(-u*u/2)
return y
def normdist(x, mu, sigma, f):
if f:
y = normcdf(x,mu,sigma)
else:
y = normpdf(x,mu,sigma)
return y
Alex's answer shows you a solution for standard normal distribution (mean = 0, standard deviation = 1). If you have normal distribution with mean and std (which is sqr(var)) and you want to calculate:
from scipy.stats import norm
# cdf(x < val)
print norm.cdf(val, m, s)
# cdf(x > val)
print 1 - norm.cdf(val, m, s)
# cdf(v1 < x < v2)
print norm.cdf(v2, m, s) - norm.cdf(v1, m, s)
Read more about cdf here and scipy implementation of normal distribution with many formulas here.
Taken from above:
from scipy.stats import norm
>>> norm.cdf(1.96)
0.9750021048517795
>>> norm.cdf(-1.96)
0.024997895148220435
For a two-tailed test:
Import numpy as np
z = 1.96
p_value = 2 * norm.cdf(-np.abs(z))
0.04999579029644087
Simple like this:
import math
def my_cdf(x):
return 0.5*(1+math.erf(x/math.sqrt(2)))
I found the formula in this page https://www.danielsoper.com/statcalc/formulas.aspx?id=55
I do have a function, for example , but this can be something else as well, like a quadratic or logarithmic function. I am only interested in the domain of . The parameters of the function (a and k in this case) are known as well.
My goal is to fit a continuous piece-wise function to this, which contains alternating segments of linear functions (i.e. sloped straight segments, each with intercept of 0) and constants (i.e. horizontal segments joining the sloped segments together). The first and last segments are both sloped. And the number of segments should be pre-selected between around 9-29 (that is 5-15 linear steps + 4-14 constant plateaus).
Formally
The input function:
The fitted piecewise function:
I am looking for the optimal resulting parameters (c,r,b) (in terms of least squares) if the segment numbers (n) are specified beforehand.
The resulting constants (c) and the breakpoints (r) should be whole natural numbers, and the slopes (b) round two decimal point values.
I have tried to do the fitting numerically using the pwlf package using a segmented constant models, and further processed the resulting constant model with some graphical intuition to "slice" the constant steps with the slopes. It works to some extent, but I am sure this is suboptimal from both fitting perspective and computational efficiency. It takes multiple minutes to generate a fitting with 8 slopes on the range of 1-50000. I am sure there must be a better way to do this.
My idea would be to instead using only numerical methods/ML, the fact that we have the algebraic form of the input function could be exploited in some way to at least to use algebraic transforms (integrals) to get to a simpler optimization problem.
import numpy as np
import matplotlib.pyplot as plt
import pwlf
# The input function
def input_func(x,k,a):
return np.power(x,1/a)*k
x = np.arange(1,5e4)
y = input_func(x, 1.8, 1.3)
plt.plot(x,y);
def pw_fit(func, x_r, no_seg, *fparams):
# working on the specified range
x = np.arange(1,x_r)
y_input = func(x, *fparams)
my_pwlf = pwlf.PiecewiseLinFit(x, y_input, degree=0)
res = my_pwlf.fit(no_seg)
yHat = my_pwlf.predict(x)
# Function values at the breakpoints
y_isec = func(res, *fparams)
# Slope values at the breakpoints
slopes = np.round(y_isec / res, decimals=2)
slopes = slopes[1:]
# For the first slope value, I use the intersection of the first constant plateau and the input function
slopes = np.insert(slopes,0,np.round(y_input[np.argwhere(np.diff(np.sign(y_input - yHat))).flatten()[0]] / np.argwhere(np.diff(np.sign(y_input - yHat))).flatten()[0], decimals=2))
plateaus = np.unique(np.round(yHat))
# If due to rounding slope values (to two decimals), there is no change in a subsequent step, I just remove those segments
to_del = np.argwhere(np.diff(slopes) == 0).flatten()
slopes = np.delete(slopes,to_del + 1)
plateaus = np.delete(plateaus,to_del)
breakpoints = [np.ceil(plateaus[0]/slopes[0])]
for idx, j in enumerate(slopes[1:-1]):
breakpoints.append(np.floor(plateaus[idx]/j))
breakpoints.append(np.ceil(plateaus[idx+1]/j))
breakpoints.append(np.floor(plateaus[-1]/slopes[-1]))
return slopes, plateaus, breakpoints
slo, plat, breaks = pw_fit(input_func, 50000, 8, 1.8, 1.3)
# The piecewise function itself
def pw_calc(x, slopes, plateaus, breaks):
x = x.astype('float')
cond_list = [x < breaks[0]]
for idx, j in enumerate(breaks[:-1]):
cond_list.append((j <= x) & (x < breaks[idx+1]))
cond_list.append(breaks[-1] <= x)
func_list = [lambda x: x * slopes[0]]
for idx, j in enumerate(slopes[1:]):
func_list.append(plateaus[idx])
func_list.append(lambda x, j=j: x * j)
return np.piecewise(x, cond_list, func_list)
y_output = pw_calc(x, slo, plat, breaks)
plt.plot(x,y,y_output);
(Not important, but I think the fitted piecewise function is not continuous as it is. Intervals should be x<=r1; r1<x<=r2; ....)
As Anatolyg has pointed out, it looks to me that in the optimal solution (for the function posted at least, and probably for any where the derivative is different from zero), the horizantal segments will collapse to a point or the minimum segment length (in this case 1).
EDIT---------------------------------------------
The behavior above could only be valid if the slopes could have an intercept. If the intercepts are zero, as posted in the question, one consideration must be taken into account: Is the initial parabolic function defined in zero or nearby? Imagine the function y=0.001 *sqrt(x-1000), then the segments defined as b*x will have a slope close to zero and will be so similar to the constant segments that the best fit will be just the line that without intercept that fits better all the function.
Provided that the function is defined in zero or nearby, you can start by approximating the curve just by linear segments (with intercepts):
divide the function domain in N intervals(equal intervals or whose size is a function of the average curvature (or second derivative) of the function along the domain).
linear fit/regression in each intervals
for each interval, if a point (or bunch of points) in the extreme of any interval is better fitted by the line of the neighbor interval than the line of its interval, this point is assigned to the neighbor interval.
Repeat from 2) until no extreme points are moved.
Linear regressions might be optimized not to calculate all the covariance matrixes from scratch on each iteration, but just adding the contributions of the moved points to the previous covariance matrixes.
Then each linear segment (LSi) is replaced by a combination of a small constant segment at the beginning (Cbi), a linear segment without intercept (Si), and another constant segment at the end (Cei). This segments are easy to calculate as Si will contain the middle point of LSi, and Cbi and Cei will have respectively the begin and end values of the segment LSi. Then the intervals of each segment has to be calculated as an intersection between lines.
With this, the constant end segment will be collinear with the constant begin segment from the next interval so they will merge, resulting in a series of constant and linear segments interleaved.
But this would be a floating point start solution. Next, you will have to apply all the roundings which will mess up quite a lot all the segments as the conditions integer intervals and linear segments without slope can be very confronting. In fact, b,c,r are not totally independent. If ci and ri+1 are known, then bi+1 is already fixed
If nothing is broken so far, the final task will be to minimize the error/cost function (I assume that it will be the integral of the error between the parabolic function and the segments). My guess is that gradients here will be quite a pain, as if you change for example one ci, all the rest of the bj and cj will have to adapt as well due to the integer intervals restriction. However, if you can generalize the derivatives between parameters ( how much do I have to adapt bi+1 if ci changes a unit), you can propagate the change of one parameter to all other parameters and have kind of a gradient. Then for each interval, you can estimate what would be the ideal parameter and averaging all intervals calculate the best gradient step. Let me illustrate this:
Assuming first that r parameters are fixed, if I change c1 by one unit, b2 changes by 0.1, c2 changes by -0.2 and b3 changes by 0.2. This would be the gradient.
Then I estimate, comparing with the parabolic curve, that c1 should increase 0.5 (to reduce the cost by 10 points), b2 should increase 0.2 (to reduce the cost by 5 points), c2 should increase 0.2 (to reduce the cost by 6 points) and b3 should increase 0.1 (to reduce the cost by 9 points).
Finally, the gradient step would be (0.5/1·10 + 0.2/0.1·5 - 0.2/(-0.2)·6 + 0.1/0.2·9)/(10 + 5 + 6 + 9)~= 0.45. Thus, c1 would increase 0.45 units, b2 would increase 0.45·0.1, and so on.
When you add the r parameters to the pot, as integer intervals do not have an proper derivative, calculation is not straightforward. However, you can consider r parameters as floating points, calculate and apply the gradient step and then apply the roundings.
We can integrate the squared error function for linear and constant pieces and let SciPy optimize it. Python 3:
import matplotlib.pyplot as plt
import numpy as np
import scipy.optimize
xl = 1
xh = 50000
a = 1.3
p = 1 / a
n = 8
def split_b_and_c(bc):
return bc[::2], bc[1::2]
def solve_for_r(b, c):
r = np.empty(2 * n)
r[0] = xl
r[1:-1:2] = c / b[:-1]
r[2::2] = c / b[1:]
r[-1] = xh
return r
def linear_residual_integral(b, x):
return (
(x ** (2 * p + 1)) / (2 * p + 1)
- 2 * b * x ** (p + 2) / (p + 2)
+ b ** 2 * x ** 3 / 3
)
def constant_residual_integral(c, x):
return x ** (2 * p + 1) / (2 * p + 1) - 2 * c * x ** (p + 1) / (p + 1) + c ** 2 * x
def squared_error(bc):
b, c = split_b_and_c(bc)
r = solve_for_r(b, c)
linear = np.sum(
linear_residual_integral(b, r[1::2]) - linear_residual_integral(b, r[::2])
)
constant = np.sum(
constant_residual_integral(c, r[2::2])
- constant_residual_integral(c, r[1:-1:2])
)
return linear + constant
def evaluate(x, b, c, r):
i = 0
while x > r[i + 1]:
i += 1
return b[i // 2] * x if i % 2 == 0 else c[i // 2]
def main():
bc0 = (xl + (xh - xl) * np.arange(1, 4 * n - 2, 2) / (4 * n - 2)) ** (
p - 1 + np.arange(2 * n - 1) % 2
)
bc = scipy.optimize.minimize(
squared_error, bc0, bounds=[(1e-06, None) for i in range(2 * n - 1)]
).x
b, c = split_b_and_c(bc)
r = solve_for_r(b, c)
X = np.linspace(xl, xh, 1000)
Y = [evaluate(x, b, c, r) for x in X]
plt.plot(X, X ** p)
plt.plot(X, Y)
plt.show()
if __name__ == "__main__":
main()
I have tried to come up with a new solution myself, based on the idea of #Amo Robb, where I have partitioned the domain, and curve fitted a dual - constant and linear - piece together (with the help of np.maximum). I have used the 1 / f(x)' as the function to designate the breakpoints, but I know this is arbitrary and does not provide a global optimum. Maybe there is some optimal function for these breakpoints. But this solution is OK for me, as it might be appropriate to have a better fit at the first segments, at the expense of the error for the later segments. (The task itself is actually a cost based retail margin calculation {supply price -> added margin}, as the retail POS software can only work with such piecewise margin function).
The answer from #David Eisenstat is correct optimal solution if the parameters are allowed to be floats. Unfortunately the POS software can not use floats. It is OK to round up c-s and r-s afterwards. But the b-s should be rounded to two decimals, as those are inputted as percents, and this constraint would ruin the optimal solution with long floats. I will try to further improve my solution with both Amo's and David's valuable input. Thank You for that!
import numpy as np
import matplotlib.pyplot as plt
from scipy.optimize import curve_fit
# The input function f(x)
def input_func(x,k,a):
return np.power(x,1/a) * k
# 1 / f(x)'
def one_per_der(x,k,a):
return a / (k * np.power(x, 1/a-1))
# 1 / f(x)' inverted
def one_per_der_inv(x,k,a):
return np.power(a / (x*k), a / (1-a))
def segment_fit(start,end,y,first_val):
b, _ = curve_fit(lambda x,b: np.maximum(first_val, b*x), np.arange(start,end), y[start-1:end-1])
b = float(np.round(b, decimals=2))
bp = np.round(first_val / b)
last_val = np.round(b * end)
return b, bp, last_val
def pw_fit(end_range, no_seg, **fparams):
y_bps = np.linspace(one_per_der(1, **fparams), one_per_der(end_range,**fparams) , no_seg+1)[1:]
x_bps = np.round(one_per_der_inv(y_bps, **fparams))
y = input_func(x, **fparams)
slopes = [np.round(float(curve_fit(lambda x,b: x * b, np.arange(1,x_bps[0]), y[:int(x_bps[0])-1])[0]), decimals = 2)]
plats = [np.round(x_bps[0] * slopes[0])]
bps = []
for i, xbp in enumerate(x_bps[1:]):
b, bp, last_val = segment_fit(int(x_bps[i]+1), int(xbp), y, plats[i])
slopes.append(b); bps.append(bp); plats.append(last_val)
breaks = sorted(list(x_bps) + bps)[:-1]
# If due to rounding slope values (to two decimals), there is no change in a subsequent step, I just remove those segments
to_del = np.argwhere(np.diff(slopes) == 0).flatten()
breaks_to_del = np.concatenate((to_del * 2, to_del * 2 + 1))
slopes = np.delete(slopes,to_del + 1)
plats = np.delete(plats[:-1],to_del)
breaks = np.delete(breaks,breaks_to_del)
return slopes, plats, breaks
def pw_calc(x, slopes, plateaus, breaks):
x = x.astype('float')
cond_list = [x < breaks[0]]
for idx, j in enumerate(breaks[:-1]):
cond_list.append((j <= x) & (x < breaks[idx+1]))
cond_list.append(breaks[-1] <= x)
func_list = [lambda x: x * slopes[0]]
for idx, j in enumerate(slopes[1:]):
func_list.append(plateaus[idx])
func_list.append(lambda x, j=j: x * j)
return np.piecewise(x, cond_list, func_list)
fparams = {'k':1.8, 'a':1.2}
end_range = 5e4
no_steps = 10
x = np.arange(1, end_range)
y = input_func(x, **fparams)
slopes, plats, breaks = pw_fit(end_range, no_steps, **fparams)
y_output = pw_calc(x, slopes, plats, breaks)
plt.plot(x,y_output,y);
I have a problem with optimization of the rejection method of generating continuous random variables. I've got a density: f(x) = 3/2 (1-x^2). Here's my code:
import random
import matplotlib.pyplot as plt
import numpy as np
import time
import scipy.stats as ss
a=0 # xmin
b=1 # xmax
m=3/2 # ymax
variables = [] #list for variables
def f(x):
return 3/2 * (1 - x**2) #probability density function
reject = 0 # number of rejections
start = time.time()
while len(variables) < 100000: #I want to generate 100 000 variables
u1 = random.uniform(a,b)
u2 = random.uniform(0,m)
if u2 <= f(u1):
variables.append(u1)
else:
reject +=1
end = time.time()
print("Time: ", end-start)
print("Rejection: ", reject)
x = np.linspace(a,b,1000)
plt.hist(variables,50, density=1)
plt.plot(x, f(x))
plt.show()
ss.probplot(variables, plot=plt)
plt.show()
My first question: Is my probability plot made properly?
And the second, what is in the title. How to optimize that method? I would like to get some advice to optimize the code. Now that code takes about 0.5 seconds and there are about 50 000 rejections. Is it possible to reduce the time and number of rejections? If it's needed I can optimize using a different method of generating variables.
My first question: Is my probability plot made properly?
No. It is made versus default normal distribution. You have to pack your function f(x) into class derived from stats.rv_continuous, make it into _pdf method, and pass it to probplot
And the second, what is in the title. How to optimise that method? Is it possible to reduce the time and number of rejections?
Sure, you have the power of NumPy vector abilities at your hands. Don't ever write explicit loops - vectoriz, vectorize and vectorize!
Look at modified code below, not a single loop, everything is done via NumPy vectors. Time went down on my computer for 100000 samples (Xeon, Win10 x64, Anaconda Python 3.7) from 0.19 to 0.003.
import numpy as np
import scipy.stats as ss
import matplotlib.pyplot as plt
import time
a = 0. # xmin
b = 1. # xmax
m = 3.0/2.0 # ymax
def f(x):
return 1.5 * (1.0 - x*x) # probability density function
start = time.time()
N = 100000
u1 = np.random.uniform(a, b, N)
u2 = np.random.uniform(0.0, m, N)
negs = np.empty(N)
negs.fill(-1)
variables = np.where(u2 <= f(u1), u1, negs) # accepted samples are positive or 0, rejected are -1
end = time.time()
accept = np.extract(variables>=0.0, variables)
reject = N - len(accept)
print("Time: ", end-start)
print("Rejection: ", reject)
x = np.linspace(a, b, 1000)
plt.hist(accept, 50, density=True)
plt.plot(x, f(x))
plt.show()
ss.probplot(accept, plot=plt) # against normal distribution
plt.show()
Concerning reducing number of rejections, you could sample with 0 rejects doing inverse method, it is cubic equation so it could work with easy
UPDATE
Here is the code to use for probplot:
class my_pdf(ss.rv_continuous):
def _pdf(self, x):
return 1.5 * (1.0 - x*x)
ss.probplot(accept, dist=my_pdf(a=a, b=b, name='my_pdf'), plot=plt)
and you should get something like
Regarding your first question, scipy.stats.probplot compares your sample against the quantiles of the normal distribution. If you'd like it to compare against the quantiles of your f(x) distribution, check out the dist parameter of probplot.
In terms of making this sampling procedure faster, avoiding loops is generally the way to go. Replacing the code between start = ... and end = ... with the following resulted in a >20x speedup for me.
n_before_accept_reject = 150000
u1 = np.random.uniform(a, b, size=n_before_accept_reject)
u2 = np.random.uniform(0, m, size=n_before_accept_reject)
variables = u1[u2 <= f(u1)]
reject = n_before_accept_reject - len(variables)
Note that this will give you approximately 100000 accepted samples each time you run it. You could raise the value of n_before_accept_reject slightly to effectively guarantee that variables will always have >100000 accepted values, and then just cap the size of variables to return exactly 100000 if necessary.
Others have spoken to the probability plotting, I'm going to address the efficiency of the rejection algorithm.
Acceptance/rejection schemes are based on m(x), a "majorizing function". A majorizing function should have two properties: 1) m(x)≥ f(x) ∀ x; and 2) m(x), when scaled to be a distribution, should be easy to generate values from.
You went with the constant function m = 3/2, which meets both requirements but does not bound f(x) very closely. Integrated from zero to one, that has an area of 3/2. Your f(x), being a valid density function, has an area of 1. Consequently, ∫f(x)) / ∫m(x)) = 1 / (3/2) = 2/3. In other words, 2/3 of the values you generate from the majorizing function are accepted, and you are rejecting 1/3 of the attempts.
You need an m(x) which provides a tighter bound for f(x). I went with a line which is tangent to f(x) at x = 1/2. With a little bit of calculus to get the slope, I derived m(x) = 15/8 - 3x/2.
This choice of m(x) has an area of 9/8, so only 1/9 of the values will be rejected. A bit more calculus yielded the inverse transform generator for x's based on this m(x) is x = (5 - sqrt(25 - 24U)) / 4, where U is a uniform(0,1) random varible.
Here's an implementation, based off your original version. I wrapped the rejection scheme in a function, and created the values with a list comprehension rather than appending to a list. As you'll see if you run this, it produces a lot fewer rejections than your original version.
import random
import matplotlib.pyplot as plt
import numpy as np
import time
import math
import scipy.stats as ss
a = 0 # xmin
b = 1 # xmax
reject = 0 # number of rejections
def f(x):
return 3.0 / 2.0 * (1.0 - x**2) #probability density function
def m(x):
return 1.875 - 1.5 * x
def generate_x():
global reject
while True:
x = (5.0 - math.sqrt(25.0 - random.uniform(0.0, 24.0))) / 4.0
u = random.uniform(0, m(x))
if u <= f(x):
return x
reject += 1
start = time.time()
variables = [generate_x() for _ in range(100000)]
end = time.time()
print("Time: ", end-start)
print("Rejection: ", reject)
x = np.linspace(a,b,1000)
plt.hist(variables,50, density=1)
plt.plot(x, f(x))
plt.show()
I want to supply a negative exponent for the scipy.stats.powerlaw routine, e.g. a=-1.5, in order to draw random samples:
"""
powerlaw.pdf(x, a) = a * x**(a-1)
"""
from scipy.stats import powerlaw
R = powerlaw.rvs(a, size=100)
Why is a > 0 required, how can I supply a negative a in order to generate the random samples, and how can I supply a normalization coefficient/transform, i.e.
PDF(x,C,a) = C * x**a
The documentation is here
http://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.powerlaw.html
Thanks!
EDIT: I should add that I'm trying to replicate IDL's RANDOMP function:
http://idlastro.gsfc.nasa.gov/ftp/pro/math/randomp.pro
A PDF, integrated over its domain, must equal one. In other words, the area under a probability density function's curve must equal one.
In [36]: import scipy.integrate as integrate
In [40]: y, err = integrate.quad(lambda x: 0.5*x**(-0.5), 0, 1)
In [41]: y
Out[41]: 0.9999999999999998 # The integral is close to 1
The powerlaw density function has a domain from 0 <= x <= 1. On this domain, the integral of x**b is finite for any b > -1. When b is smaller, x**b blows up too rapidly near x = 0. So it is not a valid probability density function when b <= -1.
In [38]: integrate.quad(lambda x: x**(-1), 0, 1)
UserWarning: The maximum number of subdivisions (50) has been achieved...
# The integral blows up
Thus for x**(a-1), a must satisfy a-1 > -1 or equivalently, a > 0.
The first constant a in a * x**(a-1) is the normalizing constant which makes the integral of a * x**(a-1) over the domain [0,1] equal to 1. So you don't get to choose this constant independent of a.
Now if you change the domain to be a measurable distance away from 0, then yes, you could define a PDF of the form C * x**a for negative a. But you'd have to state what domain you want, and I don't think there is (yet) a PDF available in scipy.stats for this.
The Python package powerlaw can do this. Consider for a>1 a power law distribution with probability density function
f(x) = c * x^(-a)
for x > x_min and f(x) = 0 otherwise. Here c is a normalization factor and is determined as
c = (a-1) * x_min^(a-1).
In the example below it is a = 1.5 and x_min = 1.0 and comparing the probability density function estimated from the random sample with the PDF from the expression above gives the expected result.
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as pl
import numpy as np
import powerlaw
a, xmin = 1.5, 1.0
N = 10000
# generates random variates of power law distribution
vrs = powerlaw.Power_Law(xmin=xmin, parameters=[a]).generate_random(N)
# plotting the PDF estimated from variates
bin_min, bin_max = np.min(vrs), np.max(vrs)
bins = 10**(np.linspace(np.log10(bin_min), np.log10(bin_max), 100))
counts, edges = np.histogram(vrs, bins, density=True)
centers = (edges[1:] + edges[:-1])/2.
# plotting the expected PDF
xs = np.linspace(bin_min, bin_max, 100000)
pl.plot(xs, [(a-1)*xmin**(a-1)*x**(-a) for x in xs], color='red')
pl.plot(centers, counts, '.')
pl.xscale('log')
pl.yscale('log')
pl.savefig('powerlaw_variates.png')
returns
If r is a uniform random deviate U(0,1), then x in the following expression is a power-law distributed random deviate:
x = xmin * (1-r) ** (-1/(alpha-1))
where xmin is the smallest (positive) value above which the power-law distribution holds, and alpha is the exponent of the distribution.
If you want to generate power-law distribution, you can use a random deviation. You just have to generate a random number between [0,1] and apply the inverse method (Wolfram). In this case, the probability density function is:
p(k) = k^(-gamma)
and y is the variable uniform between 0 and 1.
y ~ U(0,1)
import numpy as np
def power_law(k_min, k_max, y, gamma):
return ((k_max**(-gamma+1) - k_min**(-gamma+1))*y + k_min**(-gamma+1.0))**(1.0/(-gamma + 1.0))
Now to generate a distribution, you just have to create an array
nodes = 1000
scale_free_distribution = np.zeros(nodes, float)
k_min = 1.0
k_max = 100*k_min
gamma = 3.0
for n in range(nodes):
scale_free_distribution[n] = power_law(k_min, k_max,np.random.uniform(0,1), gamma)
This will work to generate a power-law distribution with gamma=3.0, if you want to fix the average of distribution, you have to study Complex Networks cause the k_min depends of k_max and the average connectivity.
My answer is almost the same as Virgil's above, with the crucial difference that that alpha is actually the negative exponent of powerlaw distribution
So, if r is a uniform random deviate U(0,1), then x in the following expression is a power-law distributed random deviate:
x = xmin * (1-r) ** (-1/(alpha-1))
where xmin is the smallest (positive) value above which the power-law distribution holds, and alpha is the negative exponent of the distribution, that is the P(x) = [constant] * x**-alpha
I need to calculate binomial confidence intervals for large set of data within a script of python. Do you know any function or library of python that can do this?
Ideally I would like to have a function like this http://statpages.org/confint.html implemented on python.
Thanks for your time.
Just noting because it hasn't been posted elsewhere here that statsmodels.stats.proportion.proportion_confint lets you get a binomial confidence interval with a variety of methods. It only does symmetric intervals, though.
I would say that R (or another stats package) would probably serve you better if you have the option. That said, if you only need the binomial confidence interval you probably don't need an entire library. Here's the function in my most naive translation from javascript.
def binP(N, p, x1, x2):
p = float(p)
q = p/(1-p)
k = 0.0
v = 1.0
s = 0.0
tot = 0.0
while(k<=N):
tot += v
if(k >= x1 and k <= x2):
s += v
if(tot > 10**30):
s = s/10**30
tot = tot/10**30
v = v/10**30
k += 1
v = v*q*(N+1-k)/k
return s/tot
def calcBin(vx, vN, vCL = 95):
'''
Calculate the exact confidence interval for a binomial proportion
Usage:
>>> calcBin(13,100)
(0.07107391357421874, 0.21204372406005856)
>>> calcBin(4,7)
(0.18405151367187494, 0.9010086059570312)
'''
vx = float(vx)
vN = float(vN)
#Set the confidence bounds
vTU = (100 - float(vCL))/2
vTL = vTU
vP = vx/vN
if(vx==0):
dl = 0.0
else:
v = vP/2
vsL = 0
vsH = vP
p = vTL/100
while((vsH-vsL) > 10**-5):
if(binP(vN, v, vx, vN) > p):
vsH = v
v = (vsL+v)/2
else:
vsL = v
v = (v+vsH)/2
dl = v
if(vx==vN):
ul = 1.0
else:
v = (1+vP)/2
vsL =vP
vsH = 1
p = vTU/100
while((vsH-vsL) > 10**-5):
if(binP(vN, v, 0, vx) < p):
vsH = v
v = (vsL+v)/2
else:
vsL = v
v = (v+vsH)/2
ul = v
return (dl, ul)
While the scipy.stats module has a method .interval() to compute the equal tails confidence, it lacks a similar method to compute the highest density interval. Here is a rough way to do it using methods found in scipy and numpy.
This solution also assumes you want to use a Beta distribution as a prior. The hyper-parameters a and b are set to 1, so that the default prior is a uniform distribution between 0 and 1.
import numpy
from scipy.stats import beta
from scipy.stats import norm
def binomial_hpdr(n, N, pct, a=1, b=1, n_pbins=1e3):
"""
Function computes the posterior mode along with the upper and lower bounds of the
**Highest Posterior Density Region**.
Parameters
----------
n: number of successes
N: sample size
pct: the size of the confidence interval (between 0 and 1)
a: the alpha hyper-parameter for the Beta distribution used as a prior (Default=1)
b: the beta hyper-parameter for the Beta distribution used as a prior (Default=1)
n_pbins: the number of bins to segment the p_range into (Default=1e3)
Returns
-------
A tuple that contains the mode as well as the lower and upper bounds of the interval
(mode, lower, upper)
"""
# fixed random variable object for posterior Beta distribution
rv = beta(n+a, N-n+b)
# determine the mode and standard deviation of the posterior
stdev = rv.stats('v')**0.5
mode = (n+a-1.)/(N+a+b-2.)
# compute the number of sigma that corresponds to this confidence
# this is used to set the rough range of possible success probabilities
n_sigma = numpy.ceil(norm.ppf( (1+pct)/2. ))+1
# set the min and max values for success probability
max_p = mode + n_sigma * stdev
if max_p > 1:
max_p = 1.
min_p = mode - n_sigma * stdev
if min_p > 1:
min_p = 1.
# make the range of success probabilities
p_range = numpy.linspace(min_p, max_p, n_pbins+1)
# construct the probability mass function over the given range
if mode > 0.5:
sf = rv.sf(p_range)
pmf = sf[:-1] - sf[1:]
else:
cdf = rv.cdf(p_range)
pmf = cdf[1:] - cdf[:-1]
# find the upper and lower bounds of the interval
sorted_idxs = numpy.argsort( pmf )[::-1]
cumsum = numpy.cumsum( numpy.sort(pmf)[::-1] )
j = numpy.argmin( numpy.abs(cumsum - pct) )
upper = p_range[ (sorted_idxs[:j+1]).max()+1 ]
lower = p_range[ (sorted_idxs[:j+1]).min() ]
return (mode, lower, upper)
Just been trying this myself. If it helps here's my solution, which takes two lines of code and seems to give equivalent results to that JS page. This is the frequentist one-sided interval, I'm calling the input argument the MLE (maximum likelihood estimate) of the binomial parameter theta. I.e. mle = number of successes/number of trials. I find the upper bound of the one sided interval. The alpha value used here is therefore double the one in the JS page for the upper limit.
from scipy.stats import binom
from scipy.optimize import bisect
def binomial_ci( mle, N, alpha=0.05 ):
"""
One sided confidence interval for a binomial test.
If after N trials we obtain mle as the proportion of those
trials that resulted in success, find c such that
P(k/N < mle; theta = c) = alpha
where k/N is the proportion of successes in the set of trials,
and theta is the success probability for each trial.
"""
to_minimise = lambda c: binom.cdf(mle*N,N,c)-alpha
return bisect(to_minimise,0,1)
To find the two sided interval, call with (1-alpha/2) and alpha/2 as arguments.
The following gives exact (Clopper-Pearson) interval for binomial distribution in a simple way.
def binomial_ci(x, n, alpha=0.05):
#x is number of successes, n is number of trials
from scipy import stats
if x==0:
c1 = 0
else:
c1 = stats.beta.interval(1-alpha, x,n-x+1)[0]
if x==n:
c2=1
else:
c2 = stats.beta.interval(1-alpha, x+1,n-x)[1]
return c1, c2
You may check the code by e.g.:
p1,p2 = binomial_ci(2,7)
from scipy import stats
assert abs(stats.binom.cdf(1,7,p1)-.975)<1E-5
assert abs(stats.binom.cdf(2,7,p2)-.025)<1E-5
assert abs(binomial_ci(0,7, alpha=.1)[0])<1E-5
assert abs((1-binomial_ci(0,7, alpha=.1)[1])**7-0.05)<1E-5
assert abs(binomial_ci(7,7, alpha=.1)[1]-1)<1E-5
assert abs((binomial_ci(7,7, alpha=.1)[0])**7-0.05)<1E-5
I used the relation between the binomial proportion confidence interval and the regularized incomplete beta function, as described here:
https://en.wikipedia.org/wiki/Binomial_proportion_confidence_interval#Clopper%E2%80%93Pearson_interval
I needed to do this as well. I was using R and wanted to learn a way to work it out for myself. I would not say it is strictly pythonic.
The docstring explains most of it. It assumes you have scipy installed.
def exact_CI(x, N, alpha=0.95):
"""
Calculate the exact confidence interval of a proportion
where there is a wide range in the sample size or the proportion.
This method avoids the assumption that data are normally distributed. The sample size
and proportion are desctibed by a beta distribution.
Parameters
----------
x: the number of cases from which the proportion is calulated as a positive integer.
N: the sample size as a positive integer.
alpha : set at 0.95 for 95% confidence intervals.
Returns
-------
The proportion with the lower and upper confidence intervals as a dict.
"""
from scipy.stats import beta
x = float(x)
N = float(N)
p = round((x/N)*100,2)
intervals = [round(i,4)*100 for i in beta.interval(alpha,x,N-x+1)]
intervals.insert(0,p)
result = {'Proportion': intervals[0], 'Lower CI': intervals[1], 'Upper CI': intervals[2]}
return result
A numpy/scipy-free way of computing the same thing using the Wilson score and an approximation to the normal cumulative density function,
import math
def binconf(p, n, c=0.95):
'''
Calculate binomial confidence interval based on the number of positive and
negative events observed.
Parameters
----------
p: int
number of positive events observed
n: int
number of negative events observed
c : optional, [0,1]
confidence percentage. e.g. 0.95 means 95% confident the probability of
success lies between the 2 returned values
Returns
-------
theta_low : float
lower bound on confidence interval
theta_high : float
upper bound on confidence interval
'''
p, n = float(p), float(n)
N = p + n
if N == 0.0: return (0.0, 1.0)
p = p / N
z = normcdfi(1 - 0.5 * (1-c))
a1 = 1.0 / (1.0 + z * z / N)
a2 = p + z * z / (2 * N)
a3 = z * math.sqrt(p * (1-p) / N + z * z / (4 * N * N))
return (a1 * (a2 - a3), a1 * (a2 + a3))
def erfi(x):
"""Approximation to inverse error function"""
a = 0.147 # MAGIC!!!
a1 = math.log(1 - x * x)
a2 = (
2.0 / (math.pi * a)
+ a1 / 2.0
)
return (
sign(x) *
math.sqrt( math.sqrt(a2 * a2 - a1 / a) - a2 )
)
def sign(x):
if x < 0: return -1
if x == 0: return 0
if x > 0: return 1
def normcdfi(p, mu=0.0, sigma2=1.0):
"""Inverse CDF of normal distribution"""
if mu == 0.0 and sigma2 == 1.0:
return math.sqrt(2) * erfi(2 * p - 1)
else:
return mu + math.sqrt(sigma2) * normcdfi(p)
Astropy provides such a function (although installing and importing astropy may be a bit excessive):
astropy.stats.binom_conf_interval
I am not an expert on statistics, but binomtest is built into SciPy and produces the same results as the accepted answer:
from scipy.stats import binomtest
binomtest(13, 100).proportion_ci()
Out[11]: ConfidenceInterval(low=0.07107304618545972, high=0.21204067708744978)
binomtest(4, 7).proportion_ci()
Out[25]: ConfidenceInterval(low=0.18405156764007, high=0.9010117215575631)
It uses Clopper-Pearson exact method by default, which matches Curt's accepted answer, which gives these values, for comparison:
Usage:
>>> calcBin(13,100)
(0.07107391357421874, 0.21204372406005856)
>>> calcBin(4,7)
(0.18405151367187494, 0.9010086059570312)
It also has options for Wilson's method, with or without continuity correction, which matches TheBamf's astropy answer:
binomtest(4, 7).proportion_ci(method='wilson')
Out[32]: ConfidenceInterval(low=0.2504583645276572, high=0.8417801447485302)
binom_conf_interval(4, 7, 0.95, interval='wilson')
Out[33]: array([0.25045836, 0.84178014])
This also matches R's binom.test and statsmodels.stats.proportion.proportion_confint, according to cxrodgers' comment:
For 30 successes in 60 trials, both R's binom.test and statsmodels.stats.proportion.proportion_confint give (.37, .63) using Klopper-Pearson.
binomtest(30, 60).proportion_ci(method='exact')
Out[34]: ConfidenceInterval(low=0.3680620319424367, high=0.6319379680575633)