How do we compute the differential $dF$ knowing $d(F^2)$?

38 Views Asked by At

Given a differentiable function $F(x)$ the differential $d(F(x))$ can be defined according to the rule

$$\int f(x) dF(x) = \int f(x) F'(x) dx$$

or in other words $dF(x) = F'(x) dx$. Intuition for this comes from thinking of $dF$ as the small perturbation of $F(x)$ as we perturb $x$ by $dx$:

$dF = \frac{dF}{dx} dx = F'(x) dx$.

Now suppose we are interested in $dG$ for $G(x) = F(x)^2$. The integral rule says

$$\int f(x) dG(x) = \int f(x) G'(x) dx = 2\int f(x) F(x)F'(x) dx =2\int f(x) F(x)\frac{dF}{dx} dx=2\int f(x) F(x) dF $$

and so $dG(x) = 2F(x) dF(x)$ or $dF(x) = \frac{dF^2(x)}{2F(x)}$

HOWEVER!

Thinking of $dF$ as a probability distribution instead leads to the following derivation:

$$dF(x) = P(\omega : x \le F(\omega) \le x + dx ) $$

$$=P(\omega : x^2 \le F^2(\omega) \le (x + dx)^2 ) $$

$$=P(\omega : x^2 \le G(\omega) \le (x + dx)^2 ) $$

$$=P(\omega : x^2 \le G(\omega) \le x^2 + 2xdx+ (dx)^2 )$$

$$=P(\omega : x^2 \le G(\omega) \le x^2 + 2xdx )$$

$$ = 2x P(\omega : x^2 \le G(\omega) \le x^2 + dx ) $$

$$ = 2x dG(x^2) $$

Hence we get $dF(x) = 2x dG(x^2)$ and so $dG(x^2) = \frac{dF(x)}{2x}$. Replace $x$ with $\sqrt x$ to get $dG(x) = \frac{dF(\sqrt x)}{2 \sqrt x}$.

In particular if $dG$ is uniform ($dx = dG(x)$) the first formula gives $dx = 2F(x) dF(x)$ and $dF(x) = \frac{dx}{2F(x)}$. In order to get $dG(x) = dx$ we need $G'(x) = 1$. Hence $G(x) = x$ and $F = \sqrt x$. So we get $dF(x) = \frac{dx}{2\sqrt x}$.

However!

If $dG$ is uniform we also have $dG(x^2) = dx$ and the second formula gives $dx = dG(x^2) = \frac{dF}{2x}$ and so $dF = 2x dx$.

These derivations cannot both be right. Which one is correct and where is the mistake?

1

There are 1 best solutions below

0
On

The problem is you are squaring the distribution but also interpreting as if you squared the random variable with the distribution.

Let $X_F$ be the random variable corresponding to distribution $F(x)$.

Let $Y_F$ be the random variable corresponding to distribution $F^2(x)$.

Then : $$dF(x) = P(\{\omega : x \leq X_F(\omega) \leq x + dx \})$$

It is not: $$dF(x) = P(\{\omega : x \leq F(\omega) \leq x + dx \})$$

Infact: $$dF(x) = \mu(\{y : F(x) \leq y \leq F(x + dx) \})$$ $$dF(x) = \mu(\{y : F(x)^2 \leq y^2 \leq F(x + dx)^2 \})$$ $$dF(x) = \mu(\{y : F(x)^2 \leq y^2 \leq (F(x)+dF(x))^2 \})$$ $$dF(x) = \mu(\{y : F(x)^2 \leq y^2 \leq F(x)^2+2F(x)dF(x) \})$$ $$dF(x) = \mu(\{y : G(x) \leq y^2 \leq G(x)+2F(x)dF(x) \})$$ $$dF(x) = \mu(\{y : G(x) \leq y^2 \leq G(x+dx) \})$$