# Probability integral transformation : Why doesn't this inverse affect inequality?

I am reading this problem from DeGroot's "Probability & Statistics" 2nd edition.

I can't understand why $$\Pr[G^{-1}[F(X)]\leq z]=\Pr[F(X)\leq G(z)]$$, as nowhere is strict monotonicity assumed. I don't get why $G^{-1}[F(X)]\leq z\iff F(X)\leq G(z)$.

• The inverse of a strictly increasing function is also strictly increasing! – kjetil b halvorsen May 27 '14 at 12:18
• Yes, but nowhere "strictly increasing" is assumed. – Silent May 27 '14 at 13:05
• OK, but if the distribution function should be constant on some interval, there is zero probability on that interval, so can just be ignored! – kjetil b halvorsen May 27 '14 at 13:07
• @kjetilbhalvorsen, will you please elaborate that with answer? I will be obliged. – Silent May 27 '14 at 13:08

You are correct that $G$ might not be strictly increasing, and thus the equivalence $G^{-1}[F(X)]\leq z\iff F(X)\leq G(z)$ cannot be deduced. However, it holds in the probabilistic sense the probability of one of these conditions being true and another false is 0.
Let us decompose the event $F(X)\leq G(z)$ into two cases depending on whether $G^{-1}[F(X)]\leq z$: $$P( F(X) \leq G(z) ) = P\left([F(X)\leq G(z)] \cap [G^{-1}[F(X)]\leq z]\right) + P\left([F(X)\leq G(z)] \cap [G^{-1}[F(X)] > z] \right).$$ Due to (not necessarily strict) monotonicity of $G$, we have $G^{-1}[F(X)]\leq z \Rightarrow F(X) \leq G(z)$. This implies that the first term of the decomposition is actually $P(G^{-1}[F(X)]\leq z)$. It remains to be shown that the second term is 0.
Note that by monotonicity of $G$, $F(X)<G(z)$ implies $G^{-1}[F(X)]\leq z$ and thus the event in question, $[F(X)\leq G(z)] \cap [G^{-1}[F(X)] > z]$, is a subset of $F(X)=G(z)$. Hence, $$P\left([F(X)\leq G(z)] \cap [G^{-1}[F(X)] > z] \right)\leq P(F(X)=G(z)).$$ But, $X$ was assumed to have a continuous d.f., for which it was shown that $F(X) \sim U(0,1)$. In particular, the probability of $F(X)$ being equal to any particular value is 0. Thus, $$P\left([F(X)\leq G(z)] \cap [G^{-1}[F(X)] > z] \right) = 0.$$ Substituting this into the second term of the first equation, we conclude that $$P\left( F(X) \leq G(z) \right) = P\left(G^{-1}[F(X)]\leq z\right).$$
• 1. How due to (not necessarily strict) monotonicity of $G$, we have $G^{-1}[F(X)]\leq z \Rightarrow F(X) \leq G(z)$? I know that for strict monotonicity that argument is correct. But, how does it follow for nonstrict case? 2. Why due to monotonicity of $G$, $F(X)<G(z)\implies G^{-1}[F(X)]\leq z$? 3. Why $G^{-1}[F(X)]\leq z$ implies that $[F(X)\leq G(z)] \cap [G^{-1}[F(X)] > z]$ is a subset of $F(X)=G(z)$? 4. How $F(X) \sim U(0,1)$? – Silent May 29 '14 at 4:50
• OK, Sir, I got why $F(X) \sim U(0,1)$, it is just because probability integral transformation. Please just solve other three doubts, please! – Silent May 30 '14 at 3:41