7

I've come across a familiar math expression in my research, and I want to formally prove the following.

$s_{n+1} - s_{n} > 0$ holds for integer $n \geq 2$, where $s_n := (1-\frac{1}{n}) - (1-\frac{1}{n})^n$.

I've plotted this using software, and it seems it's true. But I've been struggling to prove this. Does anyone have clues?

Thanks.


Edited: I'm adding information as to roughly map out how I've tried. Please see below.

First, based on basic calculus.

I tried to think about functions rather than sequences, and show that both functions increase, but the former at a higher rate. To elaborate more,

$f(x) := 1 - \frac{1}{x}$ and $g(x) := (1 - \frac{1}{x})^x$

Showing $f(x)$ is increasing is easy. Took the derivative of it, and it is always positive for all $x$. The value of $f(x)$ is $\frac{1}{2}$ and its slope $f'(x)$ is $\frac{1}{4}$ at $x=2$.

What remains is to show (1) the value of $g(x)$ is less than or equal to $f(0.5) = 0.5$; (2) the slope of $g(x)$ is less than or equal to $f'(0.5) = 0.25$, and (3) the rate at which the slope increases is less than $f''(x)$. $g(0.5) = 0.25 < f(0.5) = 0.5$ (1) is done. Now (2) and (3) to go.

Showing $g(x)$ is increasing is hard. Took the logarithm of it, and tried to take the derivative of it, but to no avail.

This is pretty much it so far.

Second, based on induction.

Now, I'm trying to prove the claim, using the standard approach, but have no real progress yet.

2 Answers2

4

If you change variable to $u=\frac1n$, then letting $n\to\infty$ is the same as letting $u\to 0^+$, and $$ (1-\tfrac1n)-(1-\tfrac1n)^n = 1-u - e^{\frac{\log(1-u)}{u}} $$ The exponent $\frac{\log(1-u)}{u}$ has a removable singularity at $u=0$; if we plug that (by dividing the series expansion of $\log(1-u)$ by $u$ term for term) we find that the right-hand-side above is nice and differentiable at $u=0$, and has derivative $\frac1{2e}-1$, which is negative.

Thus, as $u$ approaches $0$ from above, the RHS will, at least eventually, approach $1-\frac1e$ from below.

(Actually, graphing the function shows that it keeps decreasing for $u=[0,\frac12]$, so the original sequence is increasing right from the beginning at $n=2$).

  • I knew the limit will converge to $1 - \frac{1}{e}$, and it was nice to see that, again!

    Hmm, I have a quick question though. In trying to convert $-(1-\frac{1}{n})^n$ into the exponential function, shouldn't the plus sign in front of it be a minus?

    – PurplePenguin Nov 20 '15 at 01:14
  • To be more specific, the first equality should be $(1-\frac{1}{n}) - (1-\frac{1}{n})^n = 1 - u - e^{\frac{\log (1-u)}{u}}$, I suppose? – PurplePenguin Nov 20 '15 at 01:18
  • @PurplePenguin: Yes, of course, typo fixed. – hmakholm left over Monica Nov 20 '15 at 01:22
  • Sorry, but I don't quite follow the last sentence. I've done a full series expansion, and I got: $1 - u - e^{-1 -\frac{1}{2} u -\frac{1}{3} u^2 - \cdots}$. Taking the derivative gives us: $ -1 - (- \frac{1}{2} - \frac{2}{3} u - \frac{3}{4} u^2 - \cdots) e^{-1 -\frac{1}{2} u -\frac{1}{3} u^2 - \cdots}$. How do we know this derivative is always negative when $u \in (0, \frac{1}{2}]$? – PurplePenguin Nov 20 '15 at 01:55
  • Nice solution. Interestingly, we can show the validity of the inequality using only Bernoulli's Inequality and straightforward arithmetic. – Mark Viola Nov 20 '15 at 04:35
  • @PurplePenguin: I think I had an idea that the constant signs in the series would show that the function is concave and therefore if the slope is still negative at $u=1/2$ it would have been negative all the way. But I can't get the details to fit now; perhaps I overlooked something. – hmakholm left over Monica Nov 20 '15 at 09:10
3

I thought it might be instructive to prove this inequality using only Bernoulli's Inequality and some straightforward arithmetic. To that end, we proceed.

Let $s_n=\left(1-\frac1n\right)-\left(1-\frac1n\right)^n$. Then, the forward first difference of $s_n$ is given by

$$s_{n+1}-s_n=\frac{1}{n(n+1)}+\left(\left(1-\frac1n\right)^n-\left(1-\frac1{n+1}\right)^{n+1}\right) \tag 1$$

Now, we can write the second term on the right-hand side of $(1)$ as

$$\begin{align} \left(1-\frac1n\right)^n-\left(1-\frac1{n+1}\right)^{n+1}&=\left(1-\frac1{n+1}\right)^{n+1}\left(\frac{\left(1-\frac1n\right)^n}{\left(1-\frac1{n+1}\right)^{n+1}}-1\right) \tag 2\\\\ &=\left(1-\frac1{n+1}\right)^{n+1}\left(\frac{1}{\left(1-\frac1{n+1}\right)}\left(1-\frac{1}{n^2}\right)^n-1\right) \tag 3\\\\ &\ge \left(1-\frac1{n+1}\right)^{n+1}\left(\frac{1}{\left(1-\frac1{n+1}\right)}\left(1-\frac{1}{n}\right)-1\right) \tag 4\\\\ &=-\frac1{n^2}\left(1-\frac1{n+1}\right)^{n+1}\tag 5\\\\ &=-\frac{1}{n(n+1)}\left(1-\frac1{n+1}\right)^{n}\tag 6\\\\ &\ge -\frac{1}{n(n+1)}\tag 7 \end{align}$$

Therefore, we have the expected inequality

$$s_{n+1}-s_n\ge 0$$


NOTES:

In arriving at $(2)$, we factor out the term $\left(1-\frac{1}{n+1}\right)^{n+1}$.

In going from $(2)$ to $(3)$, we noted that $\frac{\left(1-\frac1n\right)^n}{\left(1-\frac1{n+1}\right)^{n+1}}=\frac{1}{\left(1-\frac1{n+1}\right)}\left(1-\frac1{n^2}\right)^n$

In arriving at $(4)$ we used Bernoulli's Inequality.

In going from $(4)$ to $(5)$ we simplified the expression in large parentheses.

In going from $(5)$ to $(6)$ we used the equality $1-\frac{1}{n+1}=\frac{n}{n+1}$

In arriving at $(7)$, we noted that $\left(1-\frac{1}{n+1}\right)^{n}\le 1$

Mark Viola
  • 179,405
  • Thanks for another proof! I was vaguely thinking Bernoulli's inequality may play a pivotal role, but didn't know how to apply it.

    Your answer is already good enough, but are you looking for a more concise proof?

    – PurplePenguin Nov 20 '15 at 04:46
  • You're welcome. My pleasure. There might be a more concise approach, but this way forward requires only basic pre-calculus tools. – Mark Viola Nov 20 '15 at 04:50
  • I've got one thing to ask. I think I can learn a lesson here. Maybe the form of the expressions may have given you a hint toward Bernoulli's inequality, but I wonder how you actually got there. The key seems to be lying around (3) and (4). Getting to (3) to apply BI to get to (4) is one thing. Why did you think in the first place (3) is needed to apply BI? Thanks. – PurplePenguin Nov 20 '15 at 04:54
  • 1
    That's actually a good question. The expression of the type $(1\pm 1/n)^n$ lend themselves well to BI when showing monotonicity. Suffice to say that it was a bit of experience in my having taken this way forward in the past. As an exercise, see if you can convince yourself that $(1+1/n)^n$ increases but $(1+1/n)^{n+1}$ decreases (both converges to $e$). – Mark Viola Nov 20 '15 at 04:57