# Boltzmann, Entropy, and KL Divergence

Right above Boltzmann’s bust at his grave is this line: \begin{align} S = k \log W. \end{align}

This is Boltzmann’s definition of entropy. Boltzmann considers this statement the greatest achievement of his life, so he asked to have it engraved on his tomb.

This definition follows directly from the multiplicity assumption in statistical mechanics. In physics textbooks, we sometimes use $\Omega$ for multiplicity instead of $W$ \begin{align} S = k_\beta \ln{\Omega}. \end{align}

In a statistical ensemble, the state space for a quantized discrete system is equivalent to the multiplicity of the statistical ensemble under discussion. The Liouville’s Theorem argues that under Hamiltonian mechanics, the points in a statistical ensemble form an incompressible liquid. By showing that flow lines in the phase space under Hamiltonian dynamics can not cross each other, we can prove that the phase volume of the ensemble stays constant. As a result, in a reversible process, the multiplicity of an ensemble is conserved.

Intuitively then, the multiplicity of a system is a product of that of its constitutes \begin{align} \Omega_{\text{system}} = \Omega_a \Omega_b …, \end{align} and this entity is uniquely described by thermodynamic variables. In the case of an ideal gas, $\Omega (V, N, E)$ is such a state function.

Now because we want this quantity to extend linearly as we increase the volume ($V$), the total energy ($E$) and the number of particles ($N$), it is the logarithm of $\Omega$ that is more suitable. This is called the extensivity requirement. This requirement allows us to treat the entropic state function as a first order homogeneous equation, and use Legendre’s transformation to derive other thermodynamic entities such as the Gibbs Free Energy, Enthalpy and so on. \begin{align} S (\lambda E, \lambda V, \lambda N) = \lambda S(E, V, N), \;\text{where}\; S = k_\beta \ln \Omega. \end{align}

#### Temperature

Now, we can derive the concept of temperature. Temperature is usually derived with the Canonical ensemble. Canonical ensemble describes a small system (usually quantized and discrete in the textbook) in thermal equilibrium with a larger thermal bath (thermostat). In addition the whole system is isolated with conserved overall energy. Now, because the multiplicity of the entire system has to be conserved, the overall phase volume of the combined system is at it’s maximum. Together with overall conservation of energy, we can show that in order for two systems to be in thermal “equal terms”, the following ratio has to equal \begin{align} \frac{1}{T} = \frac{\Delta{S_0}}{\Delta{E_0}} = \frac{\Delta{S_{bath}}}{\Delta{E_{bath}}}. \end{align} This, is the statistical mechanics’ definition of temperature. You can think of this definition as temperature being the constant that relates the marginal ratio between the energy labels for each macrostates versus the phase space that the macrostate occupies. For two systems to be in thermal equilibrium, both the multiplicity change and the energy transfer need to be conserved. This ratio that the two system shares is their temperature.

Note: the maximum phase volume assumption is essentially a Maximum Likelihood Estimation (MLE). You can similarly derive the Boltzmann distribution for a discrete system from this assumption alone.

Interestingly, by switching the sign of energy you can define negative temperature. Systems with bounded energy levels can have negative temperature, but ones without a bound can not. For these bounded systems (like in an Ising model), the system temperature can be swept up for $T \gt 0$ and down for $T \lt 0$. The temperature cannot cross zero though, and that is the third law of thermodynamics :)

#### Shannon’s Entropy

Now we need to derive Shannon’s entropy \begin{align} H = -\sum_{i}{P_i \cdot \log{P_i}}. \label{def_H} \end{align} Note the missing Boltzmann Constant in this defintion. The thermodynamic definition differs by a factor $k_\beta$.

Note: In thermodynamics, Boltsmann’s constant is an important factor not only in grounding the entropy in real units, but also making sure that terms are at the correct scale. $k_\beta$ is roughly $1/N_A$ where $N_A$ is the Avogadro’s number. Thus in the thermodynamic limit where the partition function has a factor of $N \;\mathrm{mole}$ in the front, $k_\beta$ keeps the terms close to 1.

To derive $\eqref{def_H}$, we need the partition function. We will again work with the Canonical ensemble. Suppose the small system in the ensemble has two microstates $1$ and $2$, then the relative probability between the occurrence of these states would be \begin{align} \frac{P_1}{P_2} = e^{-\Delta{E}/k_\beta T}. \label{rel_P} \end{align} The minus sign comes from the following: we assume that the entropy of the two microstates are both 0 because there is no disorder. Now because the overall combined system need to have its multiplicity conserved, the likelihood of $1$ happening over $2$ is one over the probability of states for the rest of the combined system.

Now $\eqref{rel_P}$ can only be true if the probability is \begin{align} P_i = Z^{-1} \cdot e^{- \beta \epsilon_i} \label{def_P} \end{align} and the partition function for the total probability \begin{align} Z = \sum_i{e^{-\beta\epsilon_i}}. \end{align}

Note that in physics, the partition function $Z$ is an important quantity from which every important thermal dynamic variable can be derived. For example the energy of the system (expected value of) is \begin{align} \langle E \rangle & = Z^{-1}\sum_i{\epsilon_i \cdot e^{- \beta \epsilon_i}} \\ & = - \frac{\partial{ \ln{Z}}}{\partial \beta}. \end{align} Similarly, we know that \begin{align} A = - k_\beta T \ln Z \end{align} where $A$ is the Helmholtz Free Energy. It is usually derived via Legengre Transformation from entropy. Now remeber \begin{align} A = U - TS, \end{align} therefore \begin{align} T S = U - A \end{align} \begin{align} S =k_\beta \ln{Z} - T^{-1} \frac{\partial{\ln{Z}}}{\partial{\beta}} \end{align} now take out a factor of $k_\beta$, and use $\eqref{def_P}$, we get Shannon’s information entropy (eq. 5.31, page 89, Sturge) \begin{align} S & = - k_\beta \ln {\big[Z^{-1} \cdot e^{-\beta E}\big]}\\ & = - k_\beta \sum{ P_i \ln{P_i} }. \end{align}

Note: $A$, $S$, $G$ (Gibbs free energy), and $H$ (enthalpy) are four main thermodynamic entities that are related by Legengre Transformation. There are 12 entities related this way but these four are the most commonly used in thermodynamics.

#### Interpretation

Equation $\eqref{def_H}$ can be rewritten as the expected value of $\ln{P_i^{-1}}$ \begin{align} H & = \langle \ln{ P^{-1} }\rangle \\ & = \sum{P_i \ln{\frac{1}{P_i}}}. \end{align} It makes sense that in a distribution, $\ln{(1/P_i)}$ is the amount of information that such event $i$ carries. Consequently, the expected value of $\ln(1/P_i)$ over all possible events ${i}$ is the information entropy of the distribution. This information is the logarithm of the volume of the ensemble in the phase space. Therefore entropy describes the amount of disorder a distribution contains.

$H(x) \ge 0$, entropy is always non-negative. $H_x = 0$ when $X$ is deterministic.

#### Joint Entropy

Between two distributions $X$ and $Y$ \begin{align} H(X,\;Y) &= - E_{x,\;y}\langle\ln P(X,Y)\rangle \\ &=\sum_{x,\,y} p(x,\,y) \frac{1}{p_{x,\,y}} \end{align}

#### Conditional Entropy

The conditional entropy of two distributions $X$ and $Y$ is \begin{align} H(X\,|\,Y) = \sum_{x,\;y}{P(x_i, y_i) \ln{\frac{P(y_i)}{P(x_i, y_i)}}}. \end{align} To derive, use Bayes’ Law.

Notice that $H(X|Y)\neq H(Y|X)$.

It is also easy to develop the chain rule $H(X, Y ) = H(X) + H(Y |X)$ and a corollary $H(X, Y |Z) = H(X|Z) + H(Y |X, Z)$

Now let’s summarize what we have so far.

• Entropy $H(X)$ is the log of state space (uncertainty) or “self-information” of a single random variable (probability distribution).
• Conditional entropy $H(X|\,Y)$ is the entropy of one random variable conditioned upon the knowledge of another.

• The reduction in uncertainty from $H(X)$ to $H(X|\,Y)$ is the mutual information \begin{align} I(X;\,Y) := H(X) - H(X|\,Y). \end{align}

• incidentally, the channel capacity $C$ of a communication channel is defined as $C = \max_{p(X)} I(X;\,Y)$

#### Relative Entropy (Kullback-Leibler Divergence)

The KL divergence is now looking very similar to the conditional entropy defined above sans a negative sign. Typically, $P$ represents the “true” distribution (data) and the $Q$ represents a model or a theory. \begin{align} KL(P\,\Vert\,Q) &= \bigg\langle\ln{\frac{P_{(x)}}{Q_{(x)}}} \bigg\rangle\\ &=\sum{P_i \ln{\frac{P_i}{Q_i}}}. \end{align} This is consistent with the meaning that $KL$ divergence is the difference between distibution $P$ on $Q$. Notice that

• $KL(P\,\Vert\,Q)$ is asymmetric (hence not a true metric)
• semi-definite: $\forall P, Q,\; KL(P\,\Vert\,Q) \ge 0.$
• linearlly additive for independent $P_i$’s and $Q_i$’s.

The semi-definitiness can be proven via Gibbs inequality.

#### Mutual Information (Interms of divergence)

The mutual information can be expressed in terms of the KL divergence \begin{align} I(X;\,Y) := KL(p(x,y)\,\Vert\,p(x)\,p(y)) \end{align} This is the information gained when moving from a model that assumes independence between $x$ and $y$ to their joint distribution.

#### Von Neumann Entropy in Quantum Systems

To adapte this classical definition of entropy to quantum mechanics, we can just use density matrix and the trace instead (via Von Neumann) \begin{align} H = - k_\beta \mathrm{Tr} (\rho \ln \rho). \end{align}