In the book Statistical inference by Casella and Berger, given that the pdf of X is $f_X(x) = \frac{2}{9}(x+1)$ for $-1 \le x \le 2$, we want to find the pdf of $Y=X^2$. But it says we are not able to directly use the theorem 2.1.8 which is

Theorem 2.1.8 Let $X$ have pdf $f_X(x)$ and let $Y=g(X)$. Suppose there exists a partition, $A_0, A_1, \dots A_k$ of $\cal{X}$ such that $P(X \in A_0)=0$ and $f_X(x)$ is continuous on each $A_i$. Further, suppose there exist functions $g_1(x), ..., g_k(x)$ defined on $A_1, ..., A_k$ respectively, satisfying

i. $g(x) = g_i(x)$ for $x\in A_i$

ii. $g_i(x)$ is monotone on $A_i$

iii. the set $\cal{Y} = \{y: y=g_i(x) \mathrm{\ for\ some\ } x \in A_i \}$ is the same for each $i=1, ..., k$

iv. $g_i^{-1}(y)$ has a continuous derivative on $\cal{Y}$ for each $i=1,...,k$.

Then, for $y \in \cal{Y}$, $$f_Y(y) = \sum^k_{i=1} f_X(g_i^{-1}(y)) \left|{d \over dy} g_i^{-1}(y) \right|$$

The $f_X(x)$ in question seems to satisfy all conditions: $f_X(x)$ is a monotonically increasing function, so we can create partitions so that $-1 \le x \le 0$ is $A_1$, $0 < x \le 2$ be $A_2$ and the rest $A_0$. Why can't we directly apply this theorem?

EDIT: it seems like the third condition does not hold. But can we find a partition that meets the third condition? Is there a way to know that whether such partition exists?

EDIT 2: By simply doing what @Dilip suggested me, I get

\begin{align} F_Y(y) &= P(Y \le y) \\ &= P(X^2 \le y) \\ &= P(-\sqrt{y} \le X \le \sqrt{y}) \\ &= P(-\sqrt{y} < X \le \sqrt{y}) \\ &= F_X(\sqrt{y}) - F_X(-\sqrt{y}) \\ f_Y(y) &= { d \over dy} F_Y(y) \\ &= {1 \over 2\sqrt{y}}f_x(\sqrt{y}) + {1 \over 2\sqrt{y}} f_X(-\sqrt{y}) \end{align} If I plug in $f_X(x)$, I get ${2 \over 9\sqrt{y}}$ which is correct only when $y \le 1$. How do I know that I need to separate the range into two subsets?

  • 1
    $\begingroup$ Instead of all the gobbledygook of the theorem, why not just apply the simple result that for $y\geq 0$, $F_Y(y)$ equals the probability of the event $X \in [-\sqrt y, \sqrt y]$ and differentiate to find $f_Y(y)$ in terms of $f_X(x)$? You will get precisely that vaunted formula, with two terms in it for $y\in [0,1]$ and one term for $y \in (1,2]$. $\endgroup$ Mar 18 '18 at 15:04
  • $\begingroup$ @DilipSarwate how do i know that $y$ must be separated into two terms $y \in [0,1]$ and $y \in (1,2]$? $\endgroup$
    – MoneyBall
    Mar 19 '18 at 1:50
  • 1
    $\begingroup$ " How do I know that I need to separate the range into two subsets?" That is exactly the case of not being able to apply the theorem directly. When 'multiple X -> single Y' then the theorem allows you to compute the pdf of Y by partitioning X into cases of 'single X_i -> single Y' but they need to be the same domain. X from -1 to 0 and X from 0 to 2 have Y from 0 to 1 and Y from 0 to 4. $\endgroup$ Mar 19 '18 at 6:17
  • 1
    $\begingroup$ How to know this case occurs the theorem does not tell. It starts after you did the work to figure this out (e.g. looking at the plot selecting subdomains of X with same image of Y) "suppose there exist a partition..." $\endgroup$ Mar 19 '18 at 6:22
  • 1
    $\begingroup$ The resolution when you "plug in $f_X(x)$" is incorrect as you forget the indicator function in$$f_X(x)=\frac{2}{9}(1+x)\mathbb{I}_{(-1,2)}(x)$$(See my own answer.) $\endgroup$
    – Xi'an
    Mar 19 '18 at 6:29

The third condition is important in that \begin{align*} \mathbb{P}(Y\in B) &= \int_B f_Y(y)\text{d}y\\ &= \mathbb{P}(g(X)\in B)\\ &= \sum_{i=1}^k \mathbb{P}(g(X)\in B,\,X\in A_i)\\ &= \sum_{i=1}^k \mathbb{P}(g_i(X)\in B,\,X\in A_i)\\ &= \sum_{i=1}^k \mathbb{P}(X\in g_i^{-1}(B)\cap A_i)\\ &= \sum_{i=1}^k \int_{g_i(g_i^{-1}(B)\cap A_i)} f_X\{g_i^{-1}(y)\}\left|{d \over dy} g_i^{-1}(y) \right|\text{d}y\\ \end{align*} and hence one needs to have $$g_i(g_i^{-1}(B)\cap A_i)=B\cap\cal{Y}$$ for $i=1,2,\ldots,k$ in order for the formula from Statistical Inference to hold. Otherwise, additional indicator functions are necessary.

The alternative suggested by Dilip Sarwate in his comment does not require the same condition: \begin{align} F_Y(y) &= \mathbb{P}(Y\le y)\\ &= \mathbb{P}(g(X)\le y)&\quad\text{[definition]}\\ &= \mathbb{P}(g(X)\in (0,y))\\ &=\sum_{i=1}^k \mathbb{P}(g(X)\in (0,y),\,X\in A_i)&\quad\text{[partition]}\\ &=\sum_{i=1}^k \mathbb{P}(g_i(X)\in (0,y),\,X\in A_i)&\quad\text{[specialisation]}\\ &=\sum_{i=1}^k \mathbb{P}(X\in g_i^{-1}\{(0,y)\},\,X\in A_i)&\quad\text{[invertibility]}\\ &=\sum_{i=1}^k \left|\int_{0}^{g_i^{-1}(y)} f_X(x)\mathbb{I}_{A_i}(x)\text{d}x\right|&\quad\text{[monotonicity]}\\ &=\sum_{i=1}^k \underbrace{\xi_i\int_{0}^{g_i^{-1}(y)} f_X(x)\mathbb{I}_{A_i}(x)\text{d}x}_{\xi_i\text{ is monotonicity of }g_i}\\ &=\int_0^y f_Y(\omega)\text{d}\omega \end{align} leading to
\begin{align}f_Y(y) &= \sum_{i=1}^k \xi_i f_X\{g_i^{-1}(y)\}\,{d \over dy} g_i^{-1}(y) \,\mathbb{I}_{A_i}\{g_i^{-1}(y)\}\\&=\sum_{i=1}^k f_X\{g_i^{-1}(y)\}\left|{d \over dy} g_i^{-1}(y) \right|\,\mathbb{I}_{A_i}\{g_i^{-1}(y)\}\end{align} In the example, $A_1=(-1,0)$ and $A_2=(0,2)$, $g_1(x)=g_2(x)=x^2$: $$f_Y(y)=\frac{2}{9}\frac{1}{2\sqrt{y}}(1-\sqrt{y})\mathbb{I}_{(-1,0)}(-\sqrt{y})+\frac{2}{9}\frac{1}{2\sqrt{y}}(1+\sqrt{y})\mathbb{I}_{(0,2)}(\sqrt{y})$$ which simplifies into $$\frac{2}{9\sqrt{y}}\mathbb{I}_{(0,1)}(y)+\frac{1}{9\sqrt{y}}(1+\sqrt{y})\mathbb{I}_{(1,4)}(y)$$

  • $\begingroup$ is it because $Y=X^2$? If so, is there a way to generalize it without assuming $Y=X^2$? $\endgroup$
    – MoneyBall
    Mar 19 '18 at 10:07
  • $\begingroup$ In the general case, replace $0$ with $-\infty$. $\endgroup$
    – Xi'an
    Mar 19 '18 at 10:13

The third condition fails.

$2^2=4$ is in the image of $\mathcal{A}_2$ but it is not in the image of $\mathcal{A}_1$.

Remark: Question has been changed.

  • $\begingroup$ the third condition is what confuses me the most. Does that mean every partition must have the same image? $\endgroup$
    – MoneyBall
    Mar 18 '18 at 7:51
  • $\begingroup$ but then if we set $A_1$ to be in $-2 \le x < 0$ and $A_2$ to be in $0 < x \le 2$, wouldn't they be the same? $\endgroup$
    – MoneyBall
    Mar 18 '18 at 7:53
  • $\begingroup$ $\mathcal{X}=\{x: f_X(x)>0\}$ and $\mathcal{A}_0, \ldots, \mathcal{A}_k$ is a partition of $\mathcal{X}$. $[-2,-1]$ can't be included. $\endgroup$ Mar 18 '18 at 8:11
  • $\begingroup$ so you're saying these partitions can only exist for $x$ such that their probabilities are greater than 0, and all partitions must have the same image. $\endgroup$
    – MoneyBall
    Mar 18 '18 at 9:06
  • $\begingroup$ yes, condition $3$ means same image. and the positive density is due to equation $(2.7.1)$. Note that these are just sufficient conditions and not necessary conditions. $\endgroup$ Mar 18 '18 at 17:10

To provide an alternative series of steps to arrive at the density of $Y$, using the "distribution function" method as @Xi'an answer does, we have

$$P(Y \leq y) = P(X^2 \leq y)$$

For the support of $X \in [-1,1]$ we have

$$P(X^2 \leq y) = P(-\sqrt{y}\leq X \leq \sqrt{y}), \;\;\; X \in [-1,1]$$

For $X\geq 1$ we have

$$P(X^2 \leq y) = \{P(X \leq \sqrt{y}),\;\;\;X\geq 1\} = P(1\leq X \leq \sqrt {y})$$


$$P(Y\leq y) = \cases {\int_{-\sqrt{y}}^{\sqrt{y}}f_X(x)dx\;\;\; |x|\leq 1 \\ \\ \int_1^{\sqrt{y}}f_X(x)dx\;\;\; x\geq 1} \\$$

Differentiating with respect to $y$ leads to @Xi'an final result. The upper branch corresponds to $y \in [0,1]$ and the lower branch to $y \in [1,4]$


Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.