What is a tight lower bound on the coupon collector time? In the classic Coupon Collector's problem, it is well known that the time $T$ necessary to complete a set of $n$ randomly-picked coupons satisfies $E[T] \sim n \ln n $,$Var(T) \sim n^2$, and $\Pr(T > n \ln n + cn) < e^{-c}$. 
This upper bound is better than the one given by the Chebyshev inequality, which would be roughly  $1/c^2$.
My question is: is there a corresponding better-than-Chebyshev lower bound for $T$? (e.g., something like $\Pr(T < n \ln n - cn) < e^{-c}$ ) ?
 A: Important Note: I've decided to remove the proof I gave originally in this answer. It was longer, more computational, used bigger hammers, and proved a weaker result as compared to the other proof I've given. All around, an inferior approach (in my view). If you're really interested, I suppose you can look at the edits.
The asymptotic results that I originally quoted and which are still found below in this answer do show that as $n \to \infty$ we can do a bit better than the bound proved in the other answer, which holds for all $n$.

The following asymptotic results hold
$$
\mathbb{P}(T > n \log n + c n ) \to 1 - e^{-e^{-c}}
$$
and
$$
\mathbb{P}(T \leq n \log n - c n ) \to e^{-e^c} \>.
$$
The constant $c \in \mathbb{R}$ and the limits are taken as $n \to \infty$. Note that, though they're separated into two results, they're pretty much the same result since $c$ is not constrained to be nonnegative in either case.
See, e.g., Motwani and Raghavan, Randomized Algorithms, pp. 60--63 for a proof.

Also: David kindly provides a proof for his stated upper bound in the comments to this answer.
A: Benjamin Doerr gives (in the chapter "
Analyzing Randomized Search Heuristics: Tools from Probability Theory" in the book "Theory of Randomized Search Heuristics", see the link for an online PDF) a somewhat simple proof of
Proposition Let $T$ be the stopping time of the coupon collection process. Then $\Pr[T\le (1-\epsilon)(n-1)\ln n]\le e^{-n^{\epsilon}}$.
This seems to give the desired asymptotics (from @cardinal's second answer), but with the advantage of being true for all $n$ and $\epsilon$.
Here is a proof sketch.
Proof Sketch: Let $X_i$ be the event that the $i$-th coupon is collected in the first $t$ draws. Thus, $\Pr[X_i=1]=(1-1/n)^t$.  The key fact is that the $X_i$ are negatively correlated, for any $I\subseteq[n]$, $\Pr[\forall i\in I, X_i=1]\le\prod_{i\in I}\Pr[X_i=1]$.  Intuitively, this is fairly clear, as knowing that the $i$-th coupon in the first $t$ draws would make it less likely that the $j$-th coupon is also drawn in the first $t$ draws.
One can prove the claim, but enlarging the set $I$ by 1 at each step.  Then it reduces to showing that $\Pr[\forall i\in I, X_i=1|X_j=1]\le\Pr[\forall i\in I,X_i=1]$, for $j\notin I$. Equivalently, by averaging, it reduces to showing that $\Pr[\forall i\in I, X_i=1|X_j=0]\ge\Pr[\forall i\in I,X_i=1]$. Doerr only gives an intuitive argument for this.  One avenue to a proof is as follows. One can observe that conditioned on the $j$ coupon coming after all of the coupons in $I$, that the probability of drawing a new coupon from $I$ after drawing $k$ so far is now $\frac{|I|-k}{n-1}$, instead of the previous $\frac{|I|-k}{n}$.  So decomposing the time to collect all coupons as a sum of geometric random variables, we can see that conditioning on the $j$-coupon coming after $I$ increases the success probabilities, and thus doing the conditioning only makes it more likely to collect the coupons earlier (by stochastic dominance: each geometric random variable is increased, in terms of stochastic dominance, by the conditioning, and this dominance can then be applied to the sum).
Given this negative correlation, it follows that $\Pr[T\le (1-\epsilon)(n-1)\ln n]\le (1-(1-1/n)^t)^n$, which gives the desired bound with $t=(1-\epsilon)(n-1)\ln n$.
Note added in proof: The link above is outdated. A new version of this result with a complete proof (that is, including the negative-correlation statement) can be found in Theorem 1.9.3 in B. Doerr. Probabilistic Tools for the Analysis of Randomized Optimization Heuristics. In B. Doerr and F. Neumann, editors, Theory of Evolutionary Computation: Recent Developments in Discrete Optimization, pages 1-87. Springer, 2020.
A: I'm providing this as a second answer since the analysis is completely elementary and provides exactly the desired result.
Proposition For $c > 0$ and $n \geq 1$,
$$
\mathbb{P}(T < n \log n - c n ) < e^{-c} \>.
$$
The idea behind the proof is simple:


*

*Represent the time until all coupons are collected as $T = \sum_{i=1}^n T_i$, where $T_i$ is the time that the $i$th (heretofore) unique coupon is collected. The $T_i$ are geometric random variables with mean times of $\frac{n}{n-i+1}$.

*Apply a version of the Chernoff bound and simplify.


Proof
For any $t$ and any $s > 0$, we have that
$$
\mathbb{P}(T < t) = \mathbb{P}( e^{-s T} > e^{-s t} ) \leq e^{s t} \mathbb{E} e^{-s T} \> . 
$$
Since $T = \sum_i T_i$ and the $T_i$ are independent, we can write
$$
\mathbb{E} e^{-s T} = \prod_{i=1}^n \mathbb{E} e^{- s T_i}
$$
Now since $T_i$ is geometric, let's say with probability of success $p_i$, then a simple calculation shows
$$
\mathbb{E} e^{-s T_i} = \frac{p_i}{e^s - 1 + p_i} .
$$
The $p_i$ for our problem are $p_1 = 1$, $p_2 = 1 - 1/n$, $p_3 = 1 - 2/n$, etc. Hence,
$$
\prod_{i=1}^n \mathbb{E} e^{-s T_i} = \prod_{i=1}^n \frac{i/n}{e^s - 1 + i/n}.
$$
Let's choose $s = 1/n$ and $t = n \log n - c n$ for some $c > 0$. Then
$$
e^{s t} = n e^{-c}
$$
and $e^s = e^{1/n} \geq 1 + 1/n$, yielding
$$
\prod_{i=1}^n \frac{i/n}{e^s - 1 + i/n} \leq \prod_{i=1}^n \frac{i}{i+1} = \frac{1}{n+1} \> .
$$
Putting this together, we get that
$$
P(T < n \log n - c n) \leq \frac{n}{n+1} e^{-c} < e^{-c}
$$
as desired.
A: Although @cardinal has already given an answer that gives precisely the bound I was looking for, I have found a similar Chernoff-style argument that can give a stronger bound:
Proposition:
$$
Pr (T \leq n \log n - c n) \leq \exp(- \frac{3c^2}{\pi^2}  ) \> .
$$
(this is stronger for $c > \frac{\pi^2}{3}$ )
Proof:
As in @cardinal's answer, we can use the fact that $T$ is a sum of independent geometric random variables $T_i$ with success probabilities $p_i = 1 - i/n$. It follows that $E[T_i] = 1/p_i$ and $E[T] = \sum_{i=1}^{n} E[T_i] = n \sum_{i=1}^n \frac{1}{i}\geq n \log n$.
Define now new variables $S_i  : = T_i  - E[T_i]$, and $S : = \sum_i S_i$. We can then write
$$
\Pr (T \leq n \log n - c n)  \leq \Pr (T \leq E[T] - c n) = \Pr (S \leq - c n)
$$
$$
= \Pr\left(\exp(-s S ) \geq \exp( s cn) \right) \leq e^{-s c n} E\left[ e^{-s S} \right]
$$
Computing the averages, we have
$$
E[e^{-s S}] = \prod_i E[e^{-s S_i}] =  \prod_i \frac{e^{s / p_i} } {1 + \frac{1}{p_i} (e^s -1)} 
    \leq e^{\frac{1}{2}s^2\sum_i p_i^{-2}}
$$
where the inequality follows from the facts that $e^s - 1\geq s$ and also $\frac{e^z}{1+z}\leq e^{\frac{1}{2}z^2}$ for $z\geq 0$. 
Thus, since $\sum_i p_i ^{-2} = n^2 \sum_{i=1}^{n-1} \frac{1}{i^2} \leq n^2 \pi^2/6$,  we can write
\begin{align*}
\Pr( T \leq n \log n - c n ) \leq e^{\frac{1}{12} (n \pi s)^2 - s c n}.
\end{align*}
Minimizing over $s>0$, we finally obtain
$$
\Pr( T \leq n\log n  -cn ) \leq e^{-\frac{3 c^2 }{\pi^2}}
$$
