Skip to content

Latest commit

 

History

History

nonParametrics

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 
 
 
 
 
 
 
title author date
非参估计作业
Lijun Wang
June 9, 2017

第一题

${Y_t}\sim U[-3,3]$,进行核密度估计

  • 选择不同的核函数进行估计,$K_1(x)$为高斯核,$K_2(x)=0.5I(\vert x\vert\le 1)$
  • 通过MSE和MISE选择窗宽
  • 画图$\sqrt{nh}(\hat f(x)-f(x))$

估计

首先生成$n$个$U(-3,3)$的随机数据。

n = 1000
set.seed(12345)
y = runif(n, -3, 3)

自己编写核函数估计的函数

mykde(y, kernel, h, reflect, y.range)

其中参数"y"是需要进行核密度估计的数据,"kernel"可以选择高斯核函数("gaussian")或者均匀核函数("uniform"),"h"为窗宽,"reflect"表示是否通过反射处理边缘数据,如果需要进行反射处理,则需要输入参数"y.range"来确定对称轴。具体实现细节如下:

mykde <- function(y, kernel = "gaussian", h = 0.1, reflect = FALSE, y.range=NULL )
{
  n = 1000
  x = seq(-3, 3, length.out = 1000)
  x.left = x[1:250]
  x.middle = x[251:750]
  x.right = x[751:1000]
  if (kernel == "gaussian")
  {
    kernel.gaussian <- function(x)
    {
      return(exp(-x^2/2)/sqrt(2*pi))
    }
    if (reflect)
    {
      fhat.left = sapply(x.left,
        function(x) sum(kernel.gaussian((y-x)/h) +
                          kernel.gaussian((2*y.range[1]-y-x)/h))/(n*h))
      fhat.middle = sapply(x.middle,
        function(x) sum(kernel.gaussian((y-x)/h))/(n*h))
      fhat.right = sapply(x.right,
        function(x) sum(kernel.gaussian((y-x)/h) +
                          kernel.gaussian((2*y.range[2]-y-x)/h))/(n*h))
      fhat = c(fhat.left, fhat.middle, fhat.right)
    }  
    else
      fhat = sapply(x, function(x) sum(kernel.gaussian((y-x)/h))/(n*h))
  }
  else if (kernel == "uniform")
  {
    kernel.uniform <- function(x)
    {
      x[abs(x) < 1] = 1
      x[abs(x) > 1] = 0
      return(0.5*x)
    }
    if (reflect)
    {
      fhat.left = sapply(x.left,
        function(x) sum(kernel.uniform((y-x)/h) +
                          kernel.uniform((2*y.range[1]-y-x)/h))/(n*h))
      fhat.middle = sapply(x.middle,
        function(x) sum(kernel.uniform((y-x)/h))/(n*h))
      fhat.right = sapply(x.right,
        function(x) sum(kernel.uniform((y-x)/h) +
                          kernel.uniform((2*y.range[2]-y-x)/h))/(n*h))
      fhat = c(fhat.left, fhat.middle, fhat.right)
    }
    else
      fhat = sapply(x, function(x) sum(kernel.uniform((y-x)/h))/(n*h))
  }
  else
    cast(paste0("Can not support ", kernel,
                "kernel, please change to gaussian kernel or uniform kernel.."))
  res = list(x = x,
             fhat = fhat)
  return(res)
}

取窗宽$h=0.395$,在考虑边缘效应和不考虑边缘效应时进行估计。对于反射的具体处理如下,若$y\in [a,b]$。则对较小的$y$(从小到大排序,位于前25%)取关于$y=a$的对称点,对较大的$y$(从小到大排序,位于后25%)取关于$y=b$的对称点,即 $$ \begin{array}{ll} f^(x)&=\frac 1{nh}\sum\limits_{t=1}^n[K(\frac{y_t-x}h)+K(\frac{2a-y_t-x}h)], \qquad x\in [a, a+\frac{1}4(b-a)]\ f^(x)&=\frac 1{nh}\sum\limits_{t=1}^n[K(\frac{y_t-x}h)+K(\frac{2b-y_t-x}h)], \qquad x\in [a+\frac{3}4(b-a), b] \end{array} $$

## 高斯核
mykde.gaussian = mykde(y, kernel = "gaussian", h = 0.395)
mykde.gaussian2 = mykde(y, kernel = "gaussian", h = 0.395,
                        reflect = TRUE, y.range = c(-3,3))
par(mfrow = c(1, 2))
plot(mykde.gaussian$x, mykde.gaussian$fhat, ylim = c(0, 1/5),
     cex.axis = .7, cex = 0.5)
segments(-3,1/6,3,1/6,col = "red")
plot(mykde.gaussian2$x, mykde.gaussian2$fhat, ylim = c(0, 1/5),
     cex.axis = .7, cex = 0.5)
segments(-3,1/6,3,1/6,col = "red")

左图是不考虑边缘效应的核密度估计,而右图是通过反射法来考虑边缘效应进行核密度估计。从高斯核密度估计的这两种图象可以看出,反射处理能够有效减轻边缘效应带来的核密度估计的影响。

## 均匀核
mykde.uniform = mykde(y, kernel = "uniform", h = 0.395)
mykde.uniform2 = mykde(y, kernel = "uniform", h = 0.395,
                       reflect = TRUE, y.range = c(-3,3))
par(mfrow = c(1, 2))
plot(mykde.uniform$x, mykde.uniform$fhat, ylim = c(0, 1/5),
     cex.axis = .7, cex = 0.5)
segments(-3,1/6,3,1/6,col = "red")
plot(mykde.uniform2$x, mykde.uniform2$fhat, ylim = c(0, 1/5),
     cex.axis = .7, cex = 0.5)
segments(-3,1/6,3,1/6,col = "red")

从这两张图象可以看出,反射处理能够有效减轻边缘效应带来的核密度估计的影响。从均匀核密度估计的这两张图象可以看出,反射处理能够有效减轻边缘效应带来的核密度估计的影响。

确定窗宽

下面选择高斯核函数,并采用反射方法处理边缘效应。

MSE

h = seq(0.01, 1, length.out = 100)
fhat = sapply(h, function(x) {res = mykde(y, h = x,
                              reflect = TRUE, y.range = c(-3,3)); res$fhat})
fhat.mse = apply(fhat, 2, function(x) (x-1/6)^2)
fhat.mse.min = apply(fhat.mse, 1, function(x) which.min(x))

以$y_{123}, y_{523}$为例,做出$E(\hat f(x)-f(x))^2$关于$h$的图象

par(mfrow = c(1,2))
plot(h, fhat.mse[123, ], ylab = "MSE", xlab = "h")
plot(h, fhat.mse[523, ], ylab = "MSE", xlab = "h")

从图象可以看出MSE的变化趋势,当然对于不同的点其变化趋势不同,对于每个点我们选择使得MSE最小的那个$h$,这样选出来的$h$如下(只显示了前20个x对应的$h$)

h[fhat.mse.min[1:20]]

MISE

fhat.mise = colSums(fhat.mse)/ncol(fhat.mse)
plot(h, fhat.mise, ylab = "MISE", xlab = "h")

从图象看起来,MISE是随着$h$的增大而降低的(在给定的$h$的范围内),这与上课老师讲的最优$h$在$O(h^{-1/5})$处取得似乎存在矛盾。但注意到当$f(x)=1/6$时 $$ Bias=\frac {h^2}2f''(x)\int_{-A}^Az^2K(z)dz+o(h^2) = o(h^2) $$ 则MSE主要依赖方差项, $$ Var(\hat f(x))\approx\frac 1{nh} \int_{-A}^AK^2(z)dz \cdot f(x) $$ 对于方差项,$h$越大,则方差越小,从而MSE(MISE)随着h的增大而降低,如果从这个角度看,图象中表现出的MISE随着$h$的增大而降低是合理的。

渐近分布

取$h=0.1$,采用高斯核函数进行核密度估计,考虑 $$ \sqrt{nh}(\hat f(x)-f(x)) $$ 的渐进分布

n = 10000
set.seed(12345)
y = runif(n, -3, 3)
hh = 0.1
res = mykde(y, kernel = "gaussian", h = hh, reflect = TRUE, y.range = c(-3,3))
fhat = res$fhat
plot(density(sqrt(n*hh)*(fhat-mean(fhat))), main = "limiting distribution")
# normal
x = seq(-5, 5, length.out = 1000)
sigmahat = 1.2
p = 1/sqrt(2*pi*sigmahat^2)*exp(-x^2/(sigmahat*2))
lines(x, p, col = "red")

图中黑线是$\sqrt{nh}(\hat f(x)-f(x))$在$n=10000, h=0.1$情形下的密度曲线,红色曲线是$\sigma=1.2, \mu=0$的正态分布密度曲线,可以看出两者近似程度还是相当高的,与极限分布为正态的结论是一致的。

第二题

$$ Y_t=m(X_t)+\varepsilon_t $$ 其中 $$ m(x)=2sin\pi x \qquad x\in[0,2] $$ $\varepsilon_t$任取,模拟时取$X_t\sim U[0,2]$

对$\hat m(x)$进行局部线性估计,计算$\hat m(x)$的MSE和MISE.

理论推导

由泰勒展开有 $$ m(x)=m(x_0)+(x-x_0)m'(x_0)+o(\vert x-x_0\vert), \qquad x\rightarrow x_0 $$

则有 $$ Y_t=m(x_t)=\alpha+\beta(x_t-x_0)+\varepsilon_t $$ 则局部最小二乘估计为 $$ (\hat \alpha(x_0), \hat\beta(x_0))=\underset{\alpha(x_0),\beta(x_0)}{arg; min};\sum\limits_{t=1}^n(y_t-\alpha-\beta(x_t-x_0))^2K(\frac{x_t-x_0}h) $$ 记

$$ X = \left( \begin{array}{ll} 1 & x_1-x_0\ 1 & x_2-x_0\ \cdot &\cdots\ 1 & x_n-x_0 \end{array} \right) \quad W = \left( \begin{array}{ccc} K(\frac{x_1-x_0}{h}) & &O\ & \ddots& \ O&&K(\frac{x_n-x_0}{h}) \end{array} \right) \qquad Y= \left( \begin{array}{c} y_1\ y_2\ \vdots\ y_n \end{array} \right) $$ 并记$\bm{\beta} =(\alpha,\beta)'$

则局部最小二乘估计写成矩阵形式为

$$ \hat {\bm\beta(x_0)} = \underset{\bm\beta(x_0)}{arg;min}; (Y-X\bm\beta)^TW(Y-X\bm\beta) $$ 即为加权最小二乘估计, $$ \hat{\bm\beta}(x_0)=(X'WX)^{-1}(XWY) $$ 且 $$ \hat m(x_0) = \alpha(x_0) $$ 在"lm()"函数中指定权重"weights=W"便可以求解。

编程求解

选择高斯噪声,产生$n$个数据

n = 1000
set.seed(1234)
xt = runif(n, 0, 2)
set.seed(4321)
et = rnorm(n)
yt = 2*sin(pi*xt) + et
#yt = xt + et

编写myloess函数得到局部最小二乘的估计结果

myloess(xt, yt, h, ngrid)

其中,"xt" 和"yt" 分别是数据点的横纵坐标,"h" 为窗宽,"n.grid"是网格剖分的格数,具体函数实现细节如下

myloess <- function(xt, yt, h, ngrid)
{
  x.grid = seq(0, 2, length.out = ngrid)
  kernel.gaussian <- function(x)
  {
      return(exp(-x^2/2)/sqrt(2*pi))
  }
  kernel.uniform <- function(x)
  {
      x[abs(x) < 1] = 1
      x[abs(x) > 1] = 0
      return(0.5*x)
  }
  weight <- function(x)
  {
    return(sapply(xt, function(xx) kernel.gaussian((xx-x)/h)))
  }
  w.grid = sapply(x.grid, weight)
  alpha.grid = numeric(ngrid)
  beta.grid = numeric(ngrid)
  y.grid = numeric(ngrid)
  for(i in 1:ngrid)
  {
    lm.x = xt - x.grid[i]
    lm.fit = lm(yt ~ lm.x, w = w.grid[,i])
    lm.coef = coef(lm.fit)
    alpha.grid[i] = lm.coef[[1]]
    beta.grid[i] = lm.coef[[2]]
    y.grid[i] = alpha.grid[i]
    #y.grid[i] = alpha.grid[i] + beta.grid[i]*(x.grid[i]-x.grid[i])
    #y.grid[i] = predict(lm.fit, data.frame(lm.x = x.grid[i]))
  }
  res = list(alpha = alpha.grid,
             beta = beta.grid,
             x = x.grid,
             y = y.grid)
  return(res)
}

选取窗宽$h=0.15, ngrid=512$进行模拟,得到下面结果

res = myloess(xt, yt, h = 0.15, ngrid = 512)

作出核密度估计拟合后的结果,与原始结果进行比较,可以看出局部线性估计可以很好地得到$y$与$x$原始的关系。

par(mfrow=c(1,2))
plot(xt, yt)
plot(res$x, res$y)
lines(res$x, 2*sin(pi*res$x), col = "red")

右图中红色线条为$y=2sin\pi x$的曲线,可以看出局部线性估计在极值处的估计偏小,即$x=0.5,1.5$处,其余地方与真实的差距不是很大。

MSE和MISE

$$ MSE(\hat m(x))=E(\hat m(x)-m(x))^2 $$ $$ MISE(\hat m(x)) = E\int R (\hat m(x)-m(x))^2dx = \frac{1}{N}\sum\limits{i=1}^N(\hat m(x_i)-m(x_i)) $$ 编写下列代码计算MSE和MISE

# MSE
mx.mse = (res$y - 2*sin(pi*res$x))^2
summary(mx.mse)
# MISE
mx.mise = mean(mx.mse)
mx.mise

计算出所有点的MSE,其四分位数分布结果如上所示,其均值即为MISE,即MISE=0.0172489。