When Entropy Runs Backwards: The Second Law as a Statement of Odds — Epoche B2
Introduction: A Law That Is Not Quite a Law Most textbooks present the second law of thermodynamics as an absolute rule: heat never flows from a cold body to a hot one, and entropy never decreases in an isolated system. Entropy is often glossed as "disorder", but that gloss is too vague to calculate with, and the precise meaning is what makes the argument below work. Boltzmann's definition is $S = k_B \ln W$, where $W$ counts the microscopic arrangements of atoms — positions and velocities — that are consistent with what we can actually observe about the system, and $k_B \approx 1.38 \times 10^{-23}\,\mathrm{J\,K^{-1}}$ is Boltzmann's constant. On this reading the second law is a counting statement: systems drift towards macroscopic conditions that can be realised in overwhelmingly more ways. Nothing in that formulation forbids the rare arrangement. The common belief in an absolute prohibition is understandable, but it is not exactly true. Modern statistical mechanics shows that the second law is a statement about overwhelming odds , not about impossibility. The purpose of this essay is to make that claim quantitative and to contrast two systems: a cup of tea and a microscopic bead in water. One remark on $k_B$ before the physics. It is not a constant of nature in the way that the speed of light is; it is a conversion factor, made necessary by the historical accident that temperature was measured in degrees before anyone knew that temperature is energy per degree of freedom. Its role in what follows is simply to make $\Sigma/k_B$ a pure number, and it is that pure number — the entropy change counted in units of $k_B$ — that decides everything. The Fluctuation Theorem The key result is the fluctuation theorem, proposed by Evans, Cohen and Morriss in 1993 and given a general derivation by Evans and Searles. Its subject is the entropy produced along a single trajectory — one particular history of the system, molecule by molecule — rather than the entropy of a state. For a large system every trajectory produces almost exactly the same entropy, so the distinction is invisible; for a small one, different histories give visibly different answers, and the entropy production $\Sigma_t$ over a time $t$ is a random variable with a spread. The theorem compares the probability of observing an entropy increase of size $A$ over that time with the probability of the reverse trajectory, an entropy decrease of the same size: $$\frac{P(\Sigma_t = -A)}{P(\Sigma_t = +A)} = e^{-A/k_B},$$ Reverse-entropy events are therefore not forbidden; they are exponentially suppressed. Everything depends on the size of $A$ measured in units of $k_B$ — and because the suppression is exponential, a change in $A$ by a factor of ten does not make reversals ten times rarer but unimaginably rarer. It is worth seeing at once that the theorem does not abolish the second law but derives it. Multiply both sides by $P(\Sigma_t = +A)$ and add up over all values of $A$: the left-hand side sums to $1$, since every trajectory has some entropy production, and the right-hand side is the average of $e^{-\Sigma_t/k_B}$. Hence $\langle e^{-\Sigma_t/k_B}\rangle = 1$ exactly. Now apply Jensen's inequality — the elementary fact that for a curve bending upwards, the average height of the curve over a set of points is at least its height at the average point — to the upward-bending function $e^{-x}$. It gives $\langle e^{-\Sigma_t/k_B}\rangle \ge e^{-\langle \Sigma_t\rangle/k_B}$, so $1 \ge e^{-\langle\Sigma_t\rangle/k_B}$, and therefore $$\langle \Sigma_t \rangle \ge 0.$$ The average entropy production can never be negative. The second law is exactly true of averages and only probabilistically true of individual histories. A Comparison of Two Regimes The whole question is thus the size of $A/k_B$, and the two regimes differ by twenty-four orders of magnitude. Consider first a cup of tea: take $200$ g of water at $80\,^\circ$C cooling to room temperature at $20\,^\circ$C. With the specific heat capacity of water, $c \approx 4180\ \mathrm{J\,kg^{-1}K^{-1}}$, the heat released is $$Q = mc\,\Delta T = 0.2 \times 4180 \times 60 \approx 5.0\times10^{4}\ \mathrm{J}.$$ That heat enters the room, which is so large that its temperature does not change, so its entropy rises by $Q/T_{\text{room}} = 50{,}160/293 \approx 171\ \mathrm{J\,K^{-1}}$ — the definition of entropy change for heat $Q$ absorbed at fixed temperature $T$. The tea's own entropy falls as it cools, and because its temperature changes during the process the contributions must be added up step by step, giving $mc\ln(T_f/T_i) = 836 \times \ln(293/353) \approx -156\ \mathrm{J\,K^{-1}}$. The net change is the sum: $$\Delta S \approx 171 - 156 = 15\ \mathrm{J\,K^{-1}}, \qquad \frac{\Delta S}{k_B} \approx \frac{15}{1.38\times10^{-23}} \approx 1\times10^{24}.$$ The magnitude is no accident: entropy changes of order $k_B$ per molecule, and $200$ g of water contains about $7\times10^{24}$ molecules, so a macroscopic process inevitably produces $\Sigma/k_B$ of order Avogadro's number. System Typical $\Delta S / k_B$ Suppression factor Reverse events? Cooling cup of tea $\sim 10^{24}$ $e^{-10^{24}}$ Effectively never Micron bead in water (∼2 s) $\sim 1$ $e^{-1} \approx 0.37$ Routinely observed For the tea, "never" is an excellent approximation, and it is worth converting the suppression into something the eye can grasp — the earlier version of this essay said the exponent "has of order twenty-four digits", which understates the case and mixes up two exponents. Written in base ten, $e^{-10^{24}}$ is $10^{-A}$ with $A = 10^{24}\log_{10}e \approx 4.8\times10^{23}$. The probability is thus a decimal point followed by some $4.8\times10^{23}$ zeros before the first significant figure. Writing those zeros out at one per second would take about $4.8\times10^{23}$ seconds — more than a million times the $4.4\times10^{17}$ seconds since the Big Bang. Even granting the system a fresh attempt every molecular colli