Estimating direction from a distribution on a circle Let there be $n$ points on a unit circle. It is known they come from "normal" distribution around particular unknown direction (i.e. sum of 2 "normal" distributions on circle - one centered at point $p$ and the other at its opposite $-p$). 
What is the best way to estimate this direction? By best I mean an algorithm that is a. analytical, b. efficient and c. simple.
 A: The standard way to solve this is to just consider each of your data points as unit vectors, then take the average of those unit vectors.  The direction of this averaged vector is the estimated direction.
There is a large literature on this topic which generally goes by the name of directional statistics. The seminal text on is Mardia and Jupp's book Directional Statistics. This field has a huge number of applications in astronomy, biology, meteorology, engineering etc. 
A: Ok, so now I will describe why Niels's estimator works so well. Take a bimodal and symmetric circular density function $f$ with modes $p$ and $-p$ (we will assume that $p$ is positive) such as the one plotted in my previous answer. Let $\Theta_1, \Theta_2, \dots, \Theta_N$ be $N$ observations drawn from $f$.
Niels's estimator first computes the complex numbers $e^{i 2 \Theta_n}$ and takes their average
$$ \bar{C} = \sum_{n=1}^{N} e^{i 2 \Theta_n} .$$
The estimate, denoted $\hat{p}$, is given by taking the complex argument of $\bar{C}$ and dividing by 2, that is
$$ \hat{p} = \frac{\angle{\bar{C}}}{2}$$
where $\angle{\bar{C}} \in [0,2\pi)$ denotes the complex argument. The next theorem describes the asymptotic properties of this estimator. I use the notation $\langle x \rangle_{\pi}$ do denote $x$ taken to its representative inside $[-\pi, \pi)$.  So, for example, $\langle 2\pi \rangle_{\pi} = 0$ and $\langle \pi + 0.1 \rangle_{\pi} = -\pi + 0.1$.

Theorem: 
  Let $\lambda$ denote the difference $\lambda = \tfrac{1}{2}\langle 2\hat{p} - 2p \rangle_{\pi}.$ 
  Then $\lambda$ converges almost surely to zero as $N \rightarrow \infty$ and the distribution of the normalised
  difference $\sqrt{N}\lambda$
  converges to the zero mean normal with
  variance
  $$ \frac{\sigma_s^2}{c} $$
  where
  $$ \sigma_s^2 = \int_{-\pi/2}^{\pi/2}\sin^2(\theta) f(\langle \theta + p \rangle_\pi) d\theta \qquad \text{and} \qquad c = \int_{-\pi/2}^{\pi/2}\cos(\theta) f(\langle \theta + p \rangle_\pi) d\theta. $$

The definition of the difference $\lambda$ might seem a little strange at first, but it is actually very natural. To see why note that $p$ and the estimate $\hat{p}$ are both in $[0,\pi)$ but, for example, if $p = 0$ and $\hat{p} = \pi - 0.01$ then the difference between these is not $\pi - 0.01$, because the two modes are actually very close to aligned in this case. The correct difference is $\lambda = \tfrac{1}{2}\langle 2(\pi-0.01) - 2 \times 0 \rangle_{\pi} = 0.01$.  
The proof of this theorem follows from a very similar argument to Theorem 6.1 (page 87) from my thesis. The original argument is due to Barry Quinn. Rather than restate the proof I'll just give you some convincing numerical evidence.
I've run some simulations for the case when the noise is a sum of two weighted von Mises circular distributions with concentration parameter $\kappa$.  So, when $\kappa$ is large the distribution is concentrated and looks something like the picture on the left below ($\kappa = 20$ in this case) and when $\kappa$ is small the distribution is quite spread out and looks something like the picture on the right below ($\kappa = 0.5$). We obviously expect the estimator to perform better when the distribution is quite concentrated ($\kappa$ is large). 
    
(source left) (source right)
Here are the results.  The plot below show the simulated variance of $\lambda$ after 5000 trials (the dots) versus the variance predicted in the theorem above for a range of values of $\kappa$ and number of observations $N$.  You can see that the theorem does a very good job of accurately predicting the performance if $\kappa$ isn't too small.
 (source) 
There is still an open question as to whether this is the best estimator (in the sense of maximally reducing the variance of $\lambda$). It would be possible to derive a Cramer-Rao bound for this estimation problem to give an idea of the best possible performance of an unbiased estimator. I suspect that this estimator performs very near the Cramer-Rao bound. So, in that sense it is close to best possible.
A: I see now that Andrei would like to know what to do when the distribution has 2 modes and is symmetric about these modes. It seems better to just give a second (more detailed) answer rather than complicate the simple answer I gave above (basically I think the idea in gowers comment above is sound, but it's a bit tricky to actually implement).
So, how do we deal with estimating the 'mean direction' of a distribution that looks something like:
     (source)
Good questions at this point are ''what is mean direction anyway?'' and specifically for the distribution above ''does a mean direction even exist?''
This has been a question I have been looking at a few months now. I'm wary of blowing my own horn a bit here, but I am going to attach a part of my thesis which I think gives satisfactory answers to these questions (I would love to give you the whole thesis, but it's not quite ready for the public to see). I suggest that there are (at least) two different, but equally reasonable and intuitive definitions of mean direction. I argue that the distribution above has no mean in a rigorously definable sense for both of these definitions. 
Given $N$ data points $\Theta_1,\dots, \Theta_N$ on a circle there exist very accurate and efficient O(N)-time algorithms to estimate both of these means if they exist. Neither algorithm will converge if used on circular data drawn from the bimodal distribution above as (according to my definition) the means do not exist.  
Still, given $N$ data points $\Theta_1,\dots, \Theta_N$ drawn from the bimodal distribution above, if what you want to do estimate one of the ''modes'' rather than the mean direction, then my gut tells me that there probably are efficient and accurate algorithms to do this, although I don't know if they exist in the literature. You could try Fishers book The Statistical Analysis of Circular Data.
A: This seems too simple to be true, but, combining some of the ideas posted earlier,
I think you could just interpret the vectors as complex numbers and take the RMS.
Squaring will turn the bimodal distribution into a unimodal one.
Then the square roots of the mean should give a good estimate of the modes of the original distribution.
