How and why does statistical mechanics work

As the title says we want to answer the question; how and why does statistical mechanics work? As we know from the most used prescription of Gibbs we calculate the phase space averages of dynamical quantities and we find that these phase averages agree very well with experiments. Clearly actual experiments are not done on a hypothetical ensemble they are done on the actual system in the laboratory and these experiments take a finite amount of time. Thus it is usually argued that actual measurements are time averages and they are equal to phase averages due to ergodicity. Aim of the present review is to show that ergodicity is not relevant for equilibrium statistical mechanics (with Tolman and Landau). We will see that the solution of the problem is in the very peculiar nature of the macroscopic observables and with the very large number of the degrees of freedom involved in macroscopic systems as first pointed out by Khinchin. Similar arguments are used by Landau based upon the approximate property of “Statistical Independence”. We review these ideas in detail and in some cases present a critique. We review the role of chaos (classical and quantum) where it is important and where it is not important. We criticise the ideas of E. T. Jaynes who says that the ergodic problem is conceptual one and is related to the very concept of ensemble itself which is a by-product of frequency theory of probability, and the ergodic problem becomes irrelevant when the probabilities of various micro-states are interpreted with Laplace-Bernoulli theory of Probability (Bayesian viewpoint). In the end we critically review various quantum approaches (quantum-statistical typicality approaches) to the foundations of statistical mechanics.

💡 Research Summary

The paper is a comprehensive review that tackles the long‑standing question “how and why does statistical mechanics work?” It begins by recalling the standard Gibbs prescription: macroscopic observables are obtained as phase‑space averages over an appropriate ensemble, and these averages agree remarkably well with laboratory measurements. Since real experiments are performed on a single system over a finite time, the traditional justification invokes the ergodic hypothesis, equating time averages with ensemble averages. The authors argue that this line of reasoning is unnecessary for equilibrium statistical mechanics.

The first major section surveys the “ergodic approach.” It defines the ergodic problem, presents the historical development from Boltzmann’s early attempts (including his 1868 introduction of an ergodic hypothesis) to the modern mathematical formulation by Birkhoff. The authors then discuss Khinchin’s resolution: macroscopic observables are “sum functions” – additive over many weakly interacting subsystems – and because of statistical independence the value of such observables is essentially constant on the energy hypersurface. Consequently, the overwhelming majority of microstates (the “typical” ones) yield the same macroscopic result, making the equality of time and ensemble averages a trivial consequence of the law of large numbers, not of strict ergodicity. The role of chaos and integrability is examined, with the conclusion that while chaos can guarantee ergodicity, it is not required for the success of statistical mechanics; even non‑integrable systems obey Khinchin‑Landau reasoning.

The second part presents the “non‑ergodic” viewpoint, focusing on the Landau‑Lifshitz framework. Landau derives the canonical ensemble without invoking any ergodic hypothesis, relying instead on the idea that a small subsystem in contact with a huge reservoir samples the reservoir’s microstates with equal a‑priori probability. The authors explain why this hypothesis is plausible for macroscopic systems and show how it leads directly to thermodynamic relations. Limitations of the Landau approach (e.g., non‑equilibrium situations, long‑range correlations) are acknowledged. The paper then critiques E. T. Jaynes’s information‑theoretic (Bayesian) interpretation, which claims that the ergodic problem is merely conceptual and disappears once probabilities are viewed as degrees of belief. The authors argue that, despite its elegance, Jaynes’s view does not address the physical mechanism that makes macroscopic observables insensitive to microscopic details; the mechanism remains the typicality of sum‑function observables.

The third major section turns to quantum foundations. It reviews the Eigenstate Thermalization Hypothesis (ETH), which posits that individual energy eigenstates of a many‑body quantum system already encode thermal expectation values, linking quantum chaos to thermalization. von Neumann’s quantum ergodic theorem (QET) and the later “normality” results of Goldstein, Lebowitz, Tumulka, and Zanghi are presented as rigorous statements that, for systems with many degrees of freedom, almost all pure states are thermodynamically typical. The authors organize recent quantum‑statistical typicality results into four categories: (i) kinematical canonical typicality (KCT), (ii) dynamical canonical typicality (DCT), (iii) kinematical normal typicality (KNT), and (iv) dynamical normal typicality (DNT). Each class formalizes how, either by the structure of the Hilbert space or by time evolution, most quantum states reproduce the predictions of the canonical ensemble. These quantum results echo Khinchin’s classical arguments, reinforcing the view that the sheer number of degrees of freedom, rather than strict dynamical mixing, underlies the success of statistical mechanics.

In the concluding section the authors summarize their central claim: ergodicity is not a prerequisite for equilibrium statistical mechanics. Instead, the essential ingredients are (1) the macroscopic observables being additive sum functions, (2) the statistical independence of subsystems, and (3) the enormous dimensionality of phase space or Hilbert space, which makes atypical microstates exponentially rare. They note that chaos, while interesting, is not essential; typicality arguments suffice. Open problems are identified, including experimental tests of quantum typicality, extensions to non‑equilibrium regimes, and a deeper synthesis of information‑theoretic and dynamical perspectives. An appendix revisits Boltzmann’s combinatorial derivation of entropy (S = k_B ln W). Overall, the review provides a clear, critical, and historically informed argument that the foundations of statistical mechanics rest on typicality and large‑N physics rather than on the strict validity of the ergodic hypothesis.