数值积分
- Limitation
- 对于 $d$ 维函数 $f$,一维情况下 $O(n^{-r})$ 收敛,则高维仅 $O(n^{-\frac{r}{s}})$ 收敛
- 不连续:$O(n^{-\frac{r}{s}})$
- Monte Carlo Method: $\int_0^1f(x)dx =\frac{1}{N}\sum_{i=1}^Nf(x_i)$
- Adavantage
- Easy to implement
- Robust
- efficient for high dimensional integrals
- Disadvantages
- Monte Carlo estimator: $F_N=\frac{b-a}{N}\sum_{i=1}^Nf(X_i)$
- sample
- $F_N=\frac{1}{N}\sum_{i=1}^N\frac{f(X_i)}{p(X_i)}$
- 最理想:$p(x)=\frac{f(x)}{\int f(x)dx}$
- 采样
- inversion
- 求 CDF
- 逆函数 CDF$^{-1}$
- 均匀采样
- rejection
- accept $U_1$ if $U_2<f(U_1)$
- 一般方法
- find $q(x)$, $p(x)<Mq(x)$
- Dart throwing $\xi < p(X)/Mq(X)$
- efficiency = area / area of rectangle
- transform
- $Y=y(X)$
- $P_y(y(x))=P_x(x)$
- $p_y(y)=(\frac{dy}{dx})^{-1}p_x(x)$
- $p_y(T(x))=|J_T(x)|^{-1}p_x(x)$
- Multidimensional sampling: sample with $p(x)$ and $p(y|x)$