将对数正态分布的拟合PDF缩放到python中的histrogram

时间:2016-01-20 07:09:44

标签: python scipy statistics

我有一个对数正态分布式设置样本,并希望对它进行拟合。然后我想将样本的直方图和拟合的PDF绘制成一个图,我想用直方图的原始缩放。

我的问题:如何直接缩放PDF以使其在直方图中可见?

以下是代码:

import numpy as np
import scipy.stats

# generate log-normal distributed set of samples
samples   = np.random.lognormal( mean=1., sigma=.4, size=10000 )

# make a fit to the samples and generate the resulting PDF
shape, loc, scale = scipy.stats.lognorm.fit( samples, floc=0 )
x_fit       = np.linspace( samples.min(), samples.max(), 100 )
samples_fit = scipy.stats.lognorm.pdf( x_fit, shape, loc=loc, scale=scale )

而且,为了更好地理解我的意思,这是图: Left: Samples. Middle: histogram and fitted PDF. Right: normalized histogram and fitted PDF

我的问题是,如果有一个参数可以轻松地将PDF缩放到直方图(我还没找到,但这并不意味着太多......),这样PDF就可以在中间看到情节?

1 个答案:

答案 0 :(得分:5)

您要求的是预期直方图的图表。

假设[a,b]是直方图的x个区间之一。随机 大小为n的样本,间隔中预期的样本数

(cdf(b) - cdf(a))*n

其中cdf(x)是累积分布函数。要绘制预期的直方图,您将计算每个仓的值。

下面的脚本显示了绘制预期直方图的一种方法 在matplotlib直方图之上。它生成了这个图:

histogram plot

import numpy as np
import scipy.stats
import matplotlib.pyplot as plt


# Generate log-normal distributed set of samples
np.random.seed(1234)
samples = np.random.lognormal(mean=1., sigma=.4, size=10000)

# Make a fit to the samples.
shape, loc, scale = scipy.stats.lognorm.fit(samples, floc=0)

# Create the histogram plot using matplotlib.  The first two values in
# the tuple returned by hist are the number of samples in each bin and
# the values of the histogram's bin edges.  counts has length num_bins,
# and edges has length num_bins + 1.
num_bins = 50
clr = '#FFE090'
counts, edges, patches = plt.hist(samples, bins=num_bins, color=clr, label='Sample histogram')

# Create an array of length num_bins containing the center of each bin.
centers = 0.5*(edges[:-1] + edges[1:])

# Compute the CDF at the edges. Then prob, the array of differences,
# is the probability of a sample being in the corresponding bin.
cdf = scipy.stats.lognorm.cdf(edges, shape, loc=loc, scale=scale)
prob = np.diff(cdf)

plt.plot(centers, samples.size*prob, 'k-', linewidth=2, label='Expected histogram')

# prob can also be approximated using the PDF at the centers multiplied
# by the width of the bin:
# p = scipy.stats.lognorm.pdf(centers, shape, loc=loc, scale=scale)
# prob = p*(edges[1] - edges[0])
# plt.plot(centers, samples.size*prob, 'r')

plt.legend()

plt.show()

注意:由于PDF是CDF的衍生物,您可以将cdf(b) - cdf(a)的近似值写为

cdf(b) - cdf(a) = pdf(m)*(b - a)

其中m是,例如,区间[a,b]的中点。然后,您提出的确切问题的答案是通过将PDF乘以样本大小和直方图区域宽度来缩放PDF。脚本中有一些注释掉的代码,显示如何使用缩放的PDF绘制预期的直方图。但由于CDF也可用于对数正态分布,因此您也可以使用它。