JC
Suppose we have a box of fixed volume \(V\) filled with Helium gas. How many particle \(N\) must the box contain before it makes sense to speak of the temperature of the system?
Claude
This is a wonderful question because it forces you to make precise what “makes sense” means. There isn’t a sharp threshold — it’s a crossover — but we can quantify it from several angles.
The formal setup
For \(N\) non-interacting helium atoms (monatomic ideal gas) in a box of volume \(V\) with total energy \(E\), the microcanonical phase space volume is
\[\Omega(E, V, N) \propto \frac{V^N E^{3N/2}}{N! \, \Gamma\!\left(\tfrac{3N}{2} + 1\right)}\]
and the Boltzmann entropy is \(S = k_B \ln \Omega\). Temperature is then defined by
\[\frac{1}{T} = \frac{\partial S}{\partial E}\bigg|_{V,N}\]
which gives the familiar \(\langle E \rangle = \tfrac{3}{2}Nk_BT\). Formally, you can compute this derivative for any \(N \geq 2\). Even for \(N = 3\), the formula spits out a number. So the question isn’t whether you can define \(T\) — it’s whether the definition is useful, i.e., whether temperature has the physical properties we expect of it.
Criterion 1: Fluctuations must be small
In the canonical ensemble, the energy variance of an ideal gas is
\[\frac{\Delta E}{\langle E \rangle} = \sqrt{\frac{2}{3N}}\]
This is the most direct measure of whether “the system has energy \(\tfrac{3}{2}Nk_BT\)” is a meaningful statement rather than a vacuous average. Some benchmarks:
At \(N \sim 10^2\), fluctuations are at the ~10% level. At \(N \sim 10^4\), temperature is quite sharp. By Avogadro’s number the relative fluctuation is \(\sim 10^{-12}\), which is why thermodynamics works so spectacularly for macroscopic systems.
Criterion 2: The Maxwell-Boltzmann distribution must emerge
If you look at the marginal distribution of a single particle’s speed in a microcanonical ensemble of \(N\) particles, you get a distribution that approaches the Maxwell-Boltzmann distribution as \(N \to \infty\). For finite \(N\), the marginal is actually a rescaled Beta distribution (or equivalently, a Student-\(t\) like distribution), and the deviation from Maxwell-Boltzmann is \(O(1/N)\).
For the empirical speed distribution of the \(N\) particles to look like the theoretical Maxwell-Boltzmann curve — so that a histogram would be recognizable — you need \(N\) large enough for the law of large numbers to produce reasonable statistics. In practice this requires at least \(N \sim 10^2\) to \(10^3\).
Criterion 3: Ensemble equivalence
The microcanonical and canonical ensembles agree in the thermodynamic limit, but differ at finite \(N\) by corrections of order \(O(1/N)\). Temperature is most naturally a canonical concept (the system is in contact with a heat bath), while an isolated box is microcanonical. For these to give the same physics — so that it doesn’t matter how you define \(T\) — you again need \(N\) large. The disagreement in thermodynamic predictions between ensembles is typically negligible for \(N \gtrsim 10^3\).
Criterion 4: The zeroth law must be meaningful
Temperature’s physical content comes from the zeroth law: if you partition your box into two subvolumes, they should independently have well-defined temperatures, and those temperatures should be equal in equilibrium. For this to work, each subsystem needs to be large enough to satisfy the above criteria. This means you need \(N\) large enough that \(N/2\) (or whatever partition you choose) still gives small fluctuations.
Criterion 5: Extensivity of entropy
For \(S\) to be properly extensive (so that the thermodynamic formalism is self-consistent), you need \(N\) large enough that Stirling’s approximation to \(\ln N!\) is accurate and that surface/boundary effects are negligible compared to bulk contributions. Stirling is already quite good by \(N \sim 10^2\).
The upshot
There’s no magic number, but the convergence to thermodynamic behavior is controlled by \(1/\sqrt{N}\):
- \(N \sim 10\): You can compute a formal temperature, but it fluctuates wildly and has little predictive power. The system doesn’t really have a temperature.
- \(N \sim 10^2 \text{–} 10^3\): Temperature starts to be a genuinely useful concept. Fluctuations are at the few-percent level, the Maxwell-Boltzmann distribution is recognizable, and the thermodynamic identities hold approximately.
- \(N \sim 10^4\): Temperature is sharp to better than 1%, ensemble equivalence holds cleanly, and you’re firmly in the thermodynamic regime.
So the honest answer is a few hundred to a few thousand particles — which, for helium at STP in a macroscopic box, is absurdly fewer than the \(\sim 10^{23}\) you actually have. The emergence of temperature is not a slow, grudging process. It kicks in remarkably fast, and then just gets exponentially sharper.
This is part of why statistical mechanics is such a powerful theory: you don’t need to be anywhere near the thermodynamic limit for the emergent description to be excellent. The \(1/\sqrt{N}\) convergence is your friend.