目录
概率密度函数(PDF: Probability Density Function)
累积分布函数(CDF: Cumulative Distribution Function)
核密度估计((kernel density estimation)
连续随机变量的概率分布特性。
在x点左侧事件发生的总和。
CDF特性:
①因为累计分布函数是计算x点左侧的点的数量,所以累计分布函数CDF是单调递增的。
②所有的CDF中,在x趋近-∞时,CDF趋近于0,当x趋近+∞时,CDF趋近于1。
③对于给定的数据集,CDF是唯一的
核密度估计(kernel density estimation,KDE)是在概率论中用来估计未知的密度函数,属于非参数检验方法之一,通过核密度估计图可以比较直观的看出数据样本本身的分布特征。
scipy中的stats.gaussian_kde可以计算高斯核函数的密度函数,而且提供了直接计算区间的累计密度函数,integrate_box_1d(low=-np.Inf, high=x)。
表示为:
,其中期望为μ,方差为
。

python画图效果及代码(包含随机数生成):

- import numpy as np
- import matplotlib.pyplot as plt
- import matplotlib.mlab as mlab
- import matplotlib.cm as cm
- import math
- import scipy.stats as stats
- plt.rcParams['font.sans-serif'] = ['SimHei'] # 用来正常显示中文标签
- plt.rcParams['axes.unicode_minus'] = False # 用来正常显示负号
-
- ################################ 正态分布 ###########################
- # 根据均值、标准差,求指定范围的正态分布概率值
- def normfun(x, mu, sigma):
- pdf = np.exp(-((x - mu)**2)/(2*sigma**2)) / (sigma * np.sqrt(2*np.pi))
- return pdf
- np.random.seed(0) ## 定义一个随机数种子
- result = np.random.normal(loc=10, scale=16, size=1000) # 均值为10,标准差为16
- ## !!!强调,以上参数中scale为标准差(方差的根号),不是方差,
- # 设定 x,y 轴,载入刚才的正态分布函数
- x = np.arange(min(result), max(result), 0.1)
- y = normfun(x, result.mean(), result.std())
- plt.plot(x, y) # 这里画出理论的正态分布概率曲线
- plt.hist(result, bins=20, rwidth=0.8, density=True) ## 柱状图
- plt.title('distribution')
- plt.xlabel('temperature')
- plt.ylabel('probability')
- plt.show()

- ################################ 累积分布函数cdf ###########################
- #计算正态概率密度函数在x处的值
- def norm_dist_prob(theta):
- y = stats.norm.pdf(theta, loc=np.mean(data), scale=np.std(data))
- return y
-
- #计算正态分布累积概率值
- def norm_dist_cdf(theta):
- y = stats.norm.cdf(theta,loc=np.mean(data), scale=np.std(data))
- return y
- ## 数据生成
- data = np.random.normal(loc=0.0, scale=10, size=1000)
-
- x = np.linspace(stats.norm.ppf(0.01,loc=np.mean(data), scale=np.std(data)),
- stats.norm.ppf(0.99,loc=np.mean(data), scale=np.std(data)), len(data)) #linspace() 函数返回指定间隔内均匀间隔数字的 ndarray。
-
- y1=norm_dist_prob(x)
- y2=norm_dist_cdf(x)
- plt.plot(x, y1,'g', label='pdf')
- plt.plot(x, y2,'r', label='cdf1')
- #或
- sns.kdeplot(data,cumulative=True, label='cdf2')
- plt.legend()

- ################################ 核密度估计 ###########################
- ## 数据生成
- data = np.random.normal(loc=0.0, scale=10, size=1000)
- ## 本程序是根据数据进行概率密度估计
- density = stats.gaussian_kde(data) #, bw_method=None, weights=[i[4] for i in data1]
- density.covariance_factor = lambda : .25 # lambda : .25
- density._compute_covariance()
- density.set_bandwidth(bw_method='silverman') ## 调用set_bandwidth 后计算的新带宽用于估计密度的后续评估。可选‘scott’, ‘silverman’
- xs = np.linspace(min(data), max(data), 200)
- fig, ax = plt.subplots()
- ax.plot(xs, density(xs), 'r')
- ax.fill_between(xs, density(xs), color="r", alpha=0.1)
- ax.hist(data, bins=30, rwidth=0.96, density =True, alpha=0.6,color = 'steelblue', edgecolor = 'w', label = 'dimensional histogram statistic ')
-
- ## 或者用seaborn
- fig, ax = plt.subplots()
- sns.distplot(data, hist=True, kde=True, rug=True, bins=20, ax=ax)
- # 通过hist和kde参数调节是否显示直方图及核密度估计(默认hist,kde均为True)
- # bins:int或list,控制直方图的划分
- # rug:控制是否生成观测数值的小细条
- # ax = sns.distplot(x, rug=True, rug_kws={"color": "g"},
- # ... kde_kws={"color": "k", "lw": 3, "label": "KDE"},
- # ... hist_kws={"histtype": "step", "linewidth": 3,
- # ... "alpha": 1, "color": "g"})fig, ax = plt.subplots()
两个相互独立的正态分布分别满足

则:





其生成及协方差椭圆的python实现如下:

- ################################ 二维正态分布 ###########################
- from matplotlib.patches import Ellipse
- def get_error_ellipse_parameters(cov, confidence=None, sigma=None):
- """Returns parameters of an ellipse which contains a specified
- amount of normally-distributed 2D data, where the data is
- characterised by its covariance matrix.
-
- Parameters
- ----------
- cov : array_like
- Input covariance matrix of shape (2,2)
- confidence : float
- Fraction of data points within ellipse. 0 < confidence < 1.
- If confidence is not given, it is calculated according to sigma.
- sigma : float
- Length of axes of the ellipse in standard deviations. If
- confidence is also given, sigma is ignored.
-
- Returns
- -------
- semi_major : float
- Length of major semiaxis of ellipse.
- semi_minor : float
- Length of minor semiaxis of ellipse.
- angle : float
- Rotation angle of ellipse in radian.
- confidence : float
- Fraction of data expected to lie within the ellipse.
- sigma : float
- Length of major and minor semiaxes in standard deviations.
- """
- cov = np.array(cov)
- if(cov.shape != (2,2)):
- raise ValueError("The covariance matrix needs to be of shape (2,2)")
- if(confidence == None and sigma == None):
- raise RuntimeError("One of confidence and sigma is needed as input argument")
- if(confidence and sigma):
- print("Argument sigma is ignored as confidence is also provided!")
-
- if(confidence == None):
- if(sigma < 0):
- raise ValueError("Sigma needs to be positive")
- #scaling = np.square(sigma)
- scaling = sigma
- confidence = stats.chi2.cdf(scaling, 2)
- if(sigma == None):
- if(confidence > 1 or confidence < 0):
- raise ValueError("Ensure that confidence lies between 0 and 1")
- scaling = stats.chi2.ppf(confidence, 2)
- #sigma = np.sqrt(scaling)
- sigma = scaling
- eigenvalues, eigenvectors = np.linalg.eig(cov)
- maxindex = np.argmax(eigenvalues)
- vx, vy = eigenvectors[:, maxindex]
- angle = np.arctan2(vy, vx)
- semi_minor, semi_major = np.sqrt(np.sort(eigenvalues) * scaling)
- print("With sigma = {:.3f}, {:.1f}% of data points lie within ellipse.".format(sigma, confidence * 100))
- return semi_major, semi_minor, angle, confidence, sigma
-
- mu = [1,2]
- cov = [[50,30],[30,50]] #sigma
- # 随机数生成
- z = stats.multivariate_normal(mu, cov)
- data_points = z.rvs(size = 5000)
-
- fig, ax = plt.subplots()
- plt.scatter(data_points[:,0], data_points[:,1], alpha = .5)
-
- # 画置信度椭圆
- confidence = 0.95
- semi_major, semi_minor, angle, confidence, sigma = get_error_ellipse_parameters(cov, confidence = confidence)
- ax.add_patch(Ellipse(mu, 2*semi_major, 2*semi_minor, 180*angle/np.pi, facecolor = 'none', edgecolor = 'red', label = 'Confidence = {:.0f}% (sigma = {:.2f})'.format(confidence * 100, sigma)))
- sigma = 1
- semi_major, semi_minor, angle, confidence, sigma, = get_error_ellipse_parameters(cov, sigma = sigma)
- ax.add_patch(Ellipse(mu, 2*semi_major, 2*semi_minor, 180*angle/np.pi, facecolor = 'none', edgecolor = 'yellow', label = 'Sigma = {:.0f} (confidence = {:.1f}%)'.format(sigma, confidence * 100)))
- plt.legend()
- plt.show()
计算马氏距离(Mahalanobis Distance)。一维马氏距离定义为:

- iv = [[1, 0.5, 0.5], [0.5, 1, 0.5], [0.5, 0.5, 1]]
- md = distance.mahalanobis([1, 0, 0], [0, 1, 0], iv)
- print(md)
- # 或
- p = np.array([1,1])
- distr = np.array([2,2])
- cov = [[1,0.2],
- [0.2,1]]
- dis = distance.mahalanobis(p, distr, cov)
- # p: 一个点
- # distr : 一个分布
- # 计算分布的协方差矩阵
- #cov = np.cov(distr, rowvar=False)
- # 选取分布中各维度均值所在点
- #avg_distri = np.average(distr, axis=0)
- print(dis)
卡方分布,也写作:
分布。服从自由度为n的卡方分布,记作
,其均值为 n,方差为2n。
若n个相互独立的随机变量ξ₁,ξ₂,...,ξn ,均服从标准正态分布N(0,1),则这n个服从标准正态分布的随机变量的平方和构成一新的随机变量,其分布规律称为卡方分布(chi-square distribution)。
直观说:如果 X1,X2,X3...X„是 n个具有标准正态分布的独立变量,那么其平方和
,满足具有n个自由度的
分布。

其中,
是Gamma函数,n为自由度,一般情况
:


- ################################ 卡方分布 ###########################
- for PDF in range(1,8):
- plt.plot(np.linspace(0,15,100),stats.chi2.pdf(np.linspace(0,15,100),df=PDF),label='k='+str(PDF))
- plt.tick_params(axis="both",which="major",labelsize=18)
- plt.axhline(y=0,color="black",linewidth=1.3,alpha=.7)
- plt.legend()

- ## 卡方分布相关计算
- # 累积分布函数
- x = stats.chi2.cdf(5.991, df=2)
- # 百分比点函数(与cdf—百分位数相反)
- a = stats.chi2.ppf(0.95, df=2)
- print(x,a)
- #生成随机数
- r = stats.chi2.rvs(df=df, size=1000)
Student's t-distribution,简称为t分布。
假设随机变量Z服从标准正态分布N(0,1),另一随机变量V服从m自由度的
分布,进一步假设Z和 V 彼此独立,则下列的数量t服从自由度为m的学生t分布:


- ################################ t分布 ###########################
- x = np.linspace( -3, 3, 100)
- plt.plot(x, stats.t.pdf(x,1), label='df=1')
- plt.plot(x, stats.t.pdf(x,2), label='df=20')
- plt.plot(x, stats.t.pdf(x,100), label = 'df=100')
- plt.plot( x[::5], stats.norm.pdf(x[::5]),'kx', label='normal')
- ## 累积分布函数cdf
- y = stats.t.cdf(x,df=100, loc=0, scale=1)
- plt.plot(x,y, label='cdf')
- plt.legend()