Estimate of the ratio $\frac{1}{\sqrt{2\pi}}e^{-x^2/2}(1-\text{erf}(x))$ (for standard normal distribution)

63 Views Asked by At

Define the probability density and cumulative probability of the standard Gaussian: $$ f(t) =\frac{1}{\sqrt{2\pi}} e^{-t^2/2}, \text{erf}(x) = \int_{-\infty}^x f(t) dt. $$

How can I prove that the following ratio satisfy the bound below? $$ \frac{1-\text{erf}(x)}{f(x)}=\frac{1}{\sqrt{2\pi}}e^{-x^2/2}(1-\text{erf}(x)) \geq x^{-1}-x^{-3}. $$

A lower bound not easy to find. It is always easy to bound fast-decaying functions like $f$ from above, for example with something like $e^{-x(t-x)},$ to deduce $\frac{1-\text{erf}(x)}{f(x)} < x^{-1},$ but functions dacaying even faster than $f$ are usually hard to work with. It looks like as if these are the first two terms of a series expansion, and after some search, I do find a Laurent series of this kind here, but the coefficients does not match. So this problem should be motivated by something else.

1

There are 1 best solutions below

2
On BEST ANSWER

In slightly different from this is Theorem 1.2.6 of Durrett, Probability: Theory and Examples, 5th edition. (This is probably elsewhere as well, but that's where I know it from.) Durrett observes that

$$ \int_x^\infty (1 - 3y^{-4}) \exp(-y^2/2) \: dy = (x^{-1} - x^{-3}) \exp(-x^2/2)$$

(which can be checked by differentiation). The left-hand side is obviously less than $\int_x^\infty e^{-y^2/2} \: dy$ and so you have

$$ \int_x^\infty e^{-y^2/2} \: dy \ge (x^{-1} - x^{-3}) \exp(-x^2/2) $$

The LHS is $1 - erf(x)$ and rearranging gives the bound you want.