对于给定的z
,m
,我的条件概率为p(z|m)
,其中选择系数的顺序为z
的整数,其限制为[0,1.5]
在m
范围内的1}}和[18:28]
将等于1。
def p(z,m):
if (m<21.25):
E = { 'ft':0.55, 'alpha': 2.99, 'z0':0.191, 'km':0.089, 'kt':0.25 }
S = { 'ft':0.39, 'alpha': 2.15, 'z0':0.121, 'km':0.093, 'kt':-0.175 }
I={ 'ft':0.06, 'alpha': 1.77, 'z0':0.045, 'km':0.096, 'kt':-0.9196 }
Evalue=E['ft']*np.exp(-1*E['kt']*(m-18))*z**E['alpha']*np.exp(-1*(z/(E['z0']+E['km']*(m-18)))**E['alpha'])
Svalue=S['ft']*np.exp(-1*S['kt']*(m-18))*z**S['alpha']*np.exp(-1*(z/(S['z0']+S['km']*(m-18)))**S['alpha'])
Ivalue=I['ft']*np.exp(-1*I['kt']*(m-18))*z**I['alpha']*np.exp(-1*(z/(I['z0']+I['km']*(m-18)))**I['alpha'])
value=Evalue+Svalue+Ivalue
elif(m>=21.25):
E = { 'ft':0.25, 'alpha': 1.957, 'z0':0.321, 'km':0.196, 'kt':0.565 }
S = { 'ft':0.61, 'alpha': 1.598, 'z0':0.291, 'km':0.167, 'kt':0.155 }
I = { 'ft':0.14, 'alpha': 0.964, 'z0':0.170, 'km':0.129, 'kt':0.1759 }
Evalue=E['ft']*np.exp(-1*E['kt']*(m-18))*z**E['alpha']*np.exp(-1*(z/(E['z0']+E['km']*(m-18)))**E['alpha'])
Svalue=S['ft']*np.exp(-1*S['kt']*(m-18))*z**S['alpha']*np.exp(-1*(z/(S['z0']+S['km']*(m-18)))**S['alpha'])
Ivalue=I['ft']*np.exp(-1*I['kt']*(m-18))*z**I['alpha']*np.exp(-1*(z/(I['z0']+I['km']*(m-18)))**I['alpha'])
value=Evalue+Svalue+Ivalue
return value
我想从这个分布中绘制一个样本,因此我在z
和m
平面上制作了一个网格点来估算累积分布,累计积分超过m
到达一个但z
上的累积积分并不能给我一个优势。我不知道为什么它不会融合到一起?!!
grid_m = np.linspace(18, 28, 1000)
grid_z = np.linspace(0, 1.5, 1000)
dz = np.diff(grid_z[:2])
# get cdf on grid, use cumtrapz
prob_zgm=np.empty((grid_z.shape[0], grid_m.shape[0]),float)
for i in range(grid_z.shape[0]):
for j in range(grid_m.shape[0]):
prob_zgm[i,j]=p(grid_z[i],grid_m[j])
pr = np.column_stack((np.zeros(prob_zgm.shape[0]),prob_zgm))
dm = np.diff(grid_m[:2])
cdf_zgm = integrate.cumtrapz(pr, dx=dm, axis=1)
cdf = integrate.cumtrapz(pr, dx=dz, axis=0)
哪种假设可能导致这种不一致或我错误地计算了什么?
更新:累积分布cdf_zgm
显示为
在其余部分,为了获得概率的倒数,这是我使用的方法:
# fix bounds of cdf_zgm
cdf_zgm[:, 0] = 0
cdf_zgm[:, -1] = 1
#Interpolate the data using a linear spline to "grid_q" samples
grid_q = np.linspace(0, 1, 200)
grid_qm = np.empty((len(grid_m), len(grid_q)), float)
for i in range(len(grid_m)):
grid_qm[i] = interpolate.interp1d(cdf_zgm[i], grid_z)(grid_q)
# build 2d interpolation for z as function of (q,m)
z_interp = interpolate.interp2d(grid_q, grid_m, grid_qm)
#sample magnitude
ng=20000
r = dist_m.rvs(ng)
rvs_u = np.random.rand(ng)
rvs_z = np.asarray([z_interp(rvs_u[i], r[i]) for i in range(len(rvs_u))]).ravel()
将CDF
的边界修改为one
是否正确?
答案 0 :(得分:0)
我不知道该代码有什么问题。但是这里有几个不同的想法:
(1)只需对数组元素求和,而不是尝试计算数值积分。这样更简单。 (对数组元素求和基本上是计算一个矩形规则近似值,事实证明,它实际上比梯形规则更准确。)
(2)不要试图一次创建一个完整的二维数组,而是编写一个函数,它只为给定的m值创建一个p(z | m)的一维切片。然后只需将这些元素相加以得到累积概率。