1

In the simple proof of the johnson lindenstrauss lemma written by Sanjoy Dasgupta, Anupam Gupta that can be found here they state the following (p.$62$):

Repeating this projection $O(n)$ times can boost the success probability to the desired constant, giving us the claimed randomized polynomial time algorithm.

My idea is to see that the success probability of a single trial is $1/n$ thus the success probability of atleast one trial out of $n$ trials is $1 - (1/n)^n$ if we then set it to be larger than $0.95$ we end up with: $0.05 > (1 - 1/n)^n$ thus $\log(0.05) > n\log(1 - 1/n)$ but i think this way of proving it is wrong.

I would really like to understand why the statement is true.

Could someone help me bring some clarity into this?

jakobhellander
  • 11

1 Answers1

1

Recall that $$1+x \leq e^x$$ for all real $x$.

If we run $cn$ independent trials, where $c$ is some constant, the probability of at least one success is

$$1-(1-P(\text{success}))^{cn}\geq 1-\left(1-\frac{1}{n}\right)^{cn}\geq1-e^{-c},$$

where the last line follows from the original inequality, letting $x=-\frac{1}{n}$.

For $c$ large enough, we can make the probability of at least one success as close to $1$ as we like, and this only requires $cn\in O(n)$ trials.

user51547
  • 1,775
  • 1
  • 9
  • 9