I have some questions about Chapter 5 from the book Numerical Methods for Stochastic Computations, by Dongbin Xiu.

Theorem 5.7: Let $Y$ be a random variable and $\mathbb{E}[Y^2]<\infty$. Let $Z$ be random variable and $\mathbb{E}[Z^{2m}]<\infty$ for all $m\geq0$ with generalized polynomial chaos basis $\{\Phi_k(Z)\}_{k=0}^\infty$, which satisfies $\mathbb{E}[\Phi_m(Z)\Phi_n(Z)]=\delta_{mn}\gamma_n$, $m,n\geq0$. Let $Y_N=\sum_{k=0}^N a_K \Phi_k(Z)$, where $a_k=\frac{1}{\gamma_k}\mathbb{E}[F_Y^{-1}(F_Z(Z))\Phi_k(Z)]$, $0\leq k\leq N$. Then $Y_N$ converges to $Y$ in probability.

Proof: Let $\tilde{Y}:=G(Z)=F_Y^{-1}(F_Z(Z))$. Then $\tilde{Y}$ and $Y$ have the same distribution, so $\mathbb{E}[\tilde{Y}^2]<\infty$. Since $Y_N$ is the orthogonal projection of $\tilde{Y}$, then $Y_N\rightarrow \tilde{Y}$ in $L^2$. This implies that $Y_N\rightarrow \tilde{Y}$ in probability. Since $Y$ and $\tilde{Y}$ have the same distribution, we conclude that $Y_N\rightarrow Y$ in probability.

Question 1: Why do we need $\mathbb{E}[Z^{2m}]<\infty$ ? On the other hand, I do not see why $Y_N\rightarrow Y$ in probability in the last reasoning.

In page 64, the author defines generalized polynomial chaos basis for random vectors $Z=(Z_1,\ldots,Z_d)$ with independent components. Let $\{\phi_k(Z_i)\}_{k=0}^\infty$ be a generalized polynomial chaos basis for $Z_i$: $\mathbb{E}[\phi_m(Z_i)\phi_n(Z_i)]=\delta_{mn}\gamma_m$. Let $\mathbb{i}=(i_1,\ldots,i_d)$ be a multi-index. Let $\Phi_{\mathbb{i}}(Z)=\phi_{i_1}(Z_1)\cdots\phi_{i_d}(Z_d)$, $0\leq|\mathbb{i}|\leq N$. Then $\{\Phi_{\mathbb{i}}(Z)\}_{\mathbb{i}}$ is the generalized polynomial chaos basis for $Z$.

Question 2: I think that each sequence $\{\phi_k(Z_i)\}_{k=0}^\infty$ should depend on $i$, as the orthogonality depends on the law of $Z_i$: $\{\phi_k^i(Z_i)\}_{k=0}^\infty$, $\Phi_{\mathbb{i}}(Z)=\phi_{i_1}^1(Z_1)\cdots \phi_{i_d}^d(Z_d)$. Is this correct?

Question 3: In Table 5.2, there is an example of graded lexicographic order for multi-indexes of dimension 4: $(0,0,0,0)$ is in position 1, $(1,0,0,0)$ is in position 2, $(0,1,0,0)$ is in position 3, etc. According to the definition of this order, $\mathbb{i}>\mathbb{j}$ if and only if $|\mathbb{i}|\geq|\mathbb{j}|$ and the first nonzero entry of $\mathbb{i}-\mathbb{j}$ is positive. For example, we have $(0,0,0,0)<(1,0,0,0)$. However, $(0,1,0,0)<(1,0,0,0)$, but in the table this order is reversed. Is this correct?

Question 4: In page 67, it is said the generalized polynomial chaos allows approximating the expectation and covariance of $f(t,Z)$, where $Z$ is a random vector and $t\in T$, $T$ being an index set. It is written $f_N(t,Z)=\sum_{|\mathbb{i}|\leq N}\hat{f}_{\mathbb{i}}(t)\Phi_{\mathbb{i}}(Z)$ as the approximation of $f(t,Z)$ in $L^2$. Then, in formula (5.33), it is written $\mathbb{E}[f(t,Z)]\approx \mathbb{E}[f_N(t,Z)]=\hat{f}_{\mathbb{0}}(t)$. But $\hat{f}_{\mathbb{0}}(t)$ is precisely $\mathbb{E}[f(t,Z)]$, so the approximation should be in fact an equality. Moreover, to compute the approximation $f_N(t,Z)$, we need the Fourier coefficients, so we need $\hat{f}_{\mathbb{0}}(t)=\mathbb{E}[f(t,Z)]$, which is what we wanted to approximate.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.