0

I am trying to prove the following lemma:

Let $|\varepsilon_i|\leq \varepsilon < \frac{1}{n} $ for all $1\leq i \leq n$ and define $\delta$ with

$$1+\delta=\prod_{i=1}^n (1+\varepsilon_i)^{\pm1}$$

then

$$|\delta| \leq \frac{n\varepsilon}{1-n\varepsilon}$$

The $\pm1$ in the exponent means that the exponent can be $1$ or $-1$ for any $i$ individually.

I am aware that this is a formula that allows approximations of the higher order terms of the product. As I understand it the product term has a leading $1$ and the rest are $\varepsilon_i$ dependant terms.

I have been looking through all my books searching for some useful product inequalities, however the exponent $\pm1$ and the sign of $\varepsilon_i$ rules most of them out, so I cannot use them. This inequality seemed the most promosing to deliver an upper bound:

  • Let $a_i \in \mathbb{R_0^+}$ with $\sum_{i=1}^n a_i \leq 1$, then $$\prod_{i=1}^n (1+a_i) \leq 1+2\sum_{i=1}^n a_i $$

However the nonfixed exponent bugs me and does not let me use it.

Are there any theorems or tricks I could use to get to my goal?

I am looking for hints, not for solutions. Thanks.

B.Swan
  • 2,469
  • Maybe try taking log of both sides and using things like $\log(1+\epsilon) \leq \epsilon$... – Jeff Nov 10 '17 at 17:15
  • 1
    Here are a couple things you could do depending on how tight a bound you'd like to have:
    1. split the product into two products, one where all exponents are 1, the other with -1 as the exponent then apply product-sum inequalities to each.
    2. consider the worst possible case using $\frac{1}{1-x} > 1+x$).
    – Mathily Nov 10 '17 at 17:22
  • Using the worst case scenario I have found that $\prod_{i=1}^n (1+\varepsilon_i)^{\pm1} \leq \prod_{i=1}^n (1-| \varepsilon_i |)^{-1}\leq \prod_{i=1}^n (1-\varepsilon)^{-1} = \frac{1}{(1-\varepsilon)^n} \leq \frac{1}{1-n\varepsilon} < \frac{n\varepsilon}{1-n\varepsilon}$ using Bernoulli in the 2nd last inequaliy. How do I make a statement about$ |\delta| $now? – B.Swan Nov 11 '17 at 11:00
  • The last estimation of course is false, since $n\varepsilon<1$. Here's the correct finish: $\frac{1}{1-n \varepsilon}=\frac{1-n \varepsilon+n \varepsilon}{1-n \varepsilon}=1+\frac{n\varepsilon}{1-n \varepsilon}$ – B.Swan Nov 11 '17 at 18:12

1 Answers1

1

Your final comments suggests that are already at the solution.

  1. Define $\gamma_n = \frac{nu}{1-nu}$ for nonnegative $n$ such that $nu < 1$.
  2. Show by direct calculation that $$\gamma_m + \gamma_n + \gamma_m \gamma_n \leq \gamma_{m+n}$$ for nonnegative $m$ and $n$ such that $(m+n)u < 1$.

  3. Write $\langle m \rangle $ for any error term of the form $1 + \theta$, where $|\theta| \leq m$.

  4. Realize that $$\langle m \rangle \langle n \rangle = \langle m + n \rangle$$ is just the previous inequality.
  5. Note that $u \leq \gamma_1$.
  6. Show $\frac{1}{1-\epsilon} = \langle 1 \rangle$ for $|\epsilon| \leq u$.

This compact notation, i.e. $\langle m \rangle$ saves you the grief of inventing sensible names and it automatically keeps track of the size of the error terms.

Carl Christian
  • 12,583
  • 1
  • 14
  • 37
  • I have made a mistake in my last inequality, I will correct it now. Thanks for your answer, I will take my time to understand the notation. – B.Swan Nov 11 '17 at 18:04