There is great thought experiment about it - while I completely don't see it as a realistic possibility, it is really great to understand "conflict" between 2nd law of thermodynamics and CPT symmetry conservation - that in fact there is none.

Specifically, general relativity doesn't distinguish between past and future light cones - so there appears question if spacetime has to be (time-)orientable?

... and it occurs that there are known solutions - using wormholes or slowly rotating black hole (

http://iopscience.iop.org/0264-9381/19/17/308/ ,

http://arxiv.org/PS_cache/gr-qc/pdf/0202/0202031v4.pdf ) which are not time orientable - there exist loops exchanging past and future light cones:

Imagine we have sent a rocket through such trajectory and it has returned ...

Inside this rocket, the astronaut shouldn't feel a difference - he could e.g. just break a mug ... but from our perspective it would be time reversed: pieces would get together into the mug.

Everything (like mugs) have tendency to get into higher entropic state (broken), but our things (mugs) came from past reason-result chains, so such state change (breaking) can only have e.g.: unbroken state toward past time direction and broken toward future.

In contrast, reason-result chain of mugs from the rocket came from our future time direction, so it can increase entropy only while breaking toward our past.

Great model to understand the asymmetry you are talking about is Kac ring - black and white balls in a ring and there are some marked positions. Now in each step rotate it by one position, changing color while going through marked positions.

Natural statistical assumption ("Stoßzahlansatz"): that total proportion of colors correspond to proportions before the marked positions, says it should tend to equal distribution of colors (maximal entropy) ... however, making two full rotations you get the initial setting - entropy has to finally start decreasing - here is

nice paper about it.

We can prove e.g.

Boltzmann H-theorem that in classical mechanics entropy always grows ... but making time symmetry of the system we would get opposite conclusion - contradiction. Such "theorems" for time symmetric evolution need to have some hidden extremely subtle assumption, called Stoßzahlansatz.

So in fact 2nd law works as well in both directions - if we know that in given moment all particles are in the left part of the tank, the safest/thermodynamical assumption for both later and earlier is entropy increase.

We need to find some separate explanation for asymmetry - if it is not written in fundamental equations, it has to be a property of a specific solution of physics we live in: being close to spatially localized Big Bang.