Entanglement and the foundations of statistical mechanics. S. Popescu, A. Short and A. Winter in Nature Phys. 2:754 (2006).  What the paper says!?

This shows that the density matrix of a subsystem is the canonical state, regardless of the state of the full system (subsystem + rest of the universe = full universe). They insist a great deal that their "canonical state" is whatever, and not compulsorily the Gibbs-Boltzmann form:

we are concerned only with the distance between the state of the system and the canonical state, and not with the precise mathematical form of this canonical state.

The origin of the idea is interesting:

Here we consider an alternative approach to the foundations of statistical mechanics, suggested to one of us by Yakir Aharonov about twenty years ago

The statement in the words of the authors:

Given a sufficiently small subsystem of the universe, almost every pure state of the universe is such that the subsystem is approximately in the canonical state Ω_S.

The underlying machinery is based on a high-dimensional geometry result:

A major component in the proof of the theorem is a mathematical result known as Levy’s lemma¹⁴ (presented in Fig. 2), which plays a similar role to the law of large numbers and governs the properties of typical points on high-dimensional hyperspheres.

This discussion near their conclusions is interesting, especially as it touches upon Bayesianism. Some excerpts (but the whole discussion is worth re-reading abundantly). Note the central role of entanglement, and indeed how this seems to connect to Fuchs' views on the matter.

One classical way to look at the problem:

Let us look back at what we have done. [...] One way of looking at it is to say that the only thing we know about the state of the universe is a global constraint such as its total energy. Thus, the way to proceed is to take a bayesian point of view and consider all states consistent with this global constraint to be equally probable. The average over all these states indeed leads to the state of any small subsystem being canonical. But the question then arises: what is the meaning of this average, when we deal with just one state? Also, these probabilities are subjective, and this raises the problem of how to argue for an objective meaning of the entropy. A formal way out is that suggested by Gibbs, to consider an ensemble of systems, but of course this does not solve the puzzle, because there is usually only one actual system.

The other classical way to look at it:

Alternatively, it was suggested that the state of the universe, as it evolves in time, can reach any of the states that are consistent with the global constraint. Thus, if we look at time averages, they are the same as the average that results from considering each state of the universe to be equally probable. To make sense of this image, assumptions of ergodicity are needed to ensure that the universe explores all the available space equally, and of course this does not solve the problem of what the state of the subsystem is at a given time.

This paper's input:

What we showed here is that these averages are not necessary. Rather, (almost) any individual state of the universe is such that any sufficiently small subsystem behaves as if the universe was in the equiprobable average state. This is due to massive entanglement between the subsystem and the rest of the universe, which is a generic feature of the vast majority of states. [...] The main message of our paper is that averages are not needed to justify the canonical state of a system in contact with the rest of the universe—almost any individual state of the universe is enough to lead to the canonical state. In effect, we propose to replace the postulate of equal a priori probabilities by the principle of apparently equal a priori probabilities, which states that as far as the system is concerned almost every state of the universe seems similar to the average.

This is related to Canonical Typicality by Goldstein et al.