$$ X = \begin{cases} 1, & \text{if $x \in [0,1]$} \\ 0, & \text{if $x \notin [0,1]$} \end{cases} \\[3ex] y = g(x) = \frac{X}{2} $$

Could you please explain why $p_y(y)=p_x(2y) \neq 0, \,on\, [0, \frac{1}{2}]$ instead of [0, 1]?

From MIT press book "Ian Goodfellow and Yoshua Bengio and Aaron Courville", chapter 3 "probability and information theory", page 72:

enter image description here

  • $\begingroup$ $X \in [0,1] \implies Y= X/2 \in [0,1/2],$ that's all. Loosely, the intuition that the authors are trying to develop is that under invertible maps, the probabilites are preserved, and not the density functions. Let $\delta x$ be very small. Note that $p(x) \delta x$ is roughly the probability that $X \in [x - \delta x, x+ \delta x]$, and so we should have $p(x) \delta x = p(g(x)) \delta g(x) \approx p(g(x)) g'(x) \delta x$ and not $p(g(x)) = p(x).$ This can be made more formal, of course. $\endgroup$ – stochasticboy321 Feb 23 '18 at 22:02

If one goes by the definition of the cumulative distribution function then the one writes:

$$F_y(\chi)=P(y<\chi)=P\left(\frac x2<\chi\right)=P(x<2\chi)=F_x(2\chi).$$

Hence, the corresponding probability density function is


There is the missing factor. It comes up when differentiating the cumulative density function.

(Note: In this form, this calculation is valid only in the case given by the OP.)


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.