3
$\begingroup$

On Wikipedia, the Gibbs measure defines the probability as:

$$ P(X=x) = \frac{1}{Z(\beta)}\exp(-\beta E(x)) $$

Now, the familiar form of the normal distribution is:

$$ P(x) = \frac{1}{\sqrt{2\pi}\sigma}\exp-{\frac{(x-\mu)^{2}}{2\sigma^2}} $$

Now, it seems that the normal distribution is a Gibbs measure and I was wondering whether I can equate some of the terms. So, for example what would be the $\beta$ term and more importantly, what is the corresponding energy term in the normal distribution. Can I say $\beta$ is $\frac{1}{2}$ and the energy term is $\frac{(x-\mu)^{2}}{\sigma^2}$ but then I do not understand how the partition function $Z$ is a function of $\beta$.

$\endgroup$
  • 2
    $\begingroup$ Luca: the $x$'s in your formulas have very different meanings. In the Gibbs measure, one possibility is that $x$ is a point in the configuration space of a system of particles, that is, in this case $x$ is a $3N$-dimensional vector containing the cartesian components of the positions and velocities of $n$ particles. Another possibility is that $x$ is a vector representing the discrete values of the spins in a network. In the normal distribution, the $x$ is a real number: one of the possible values of a random variable whose support is the real line. $\endgroup$ – Zen Jan 23 '14 at 19:10
  • $\begingroup$ ahhhh…so, they are not exactly comparable. How does one then use energy functions (like used for regularisation) in a probabilistic framework? $\endgroup$ – Luca Jan 23 '14 at 19:12
  • 1
    $\begingroup$ Take a look at this math.ntnu.no/~joeid/TMA4250/image_ana.pdf $\endgroup$ – Zen Jan 23 '14 at 19:35
  • 1
    $\begingroup$ Also math.uni-bremen.de/zetem/DFG-Schwerpunkt/jahrestreffen02/… $\endgroup$ – Zen Jan 23 '14 at 19:35
  • 1
    $\begingroup$ The reason for the appearence of the Gibbs measure in these problems is a deep mathematical result known as the Hammersley-Clifford Theorem. $\endgroup$ – Zen Jan 23 '14 at 19:36
2
$\begingroup$

I will not provide a full or mathematically rigorous answer, but I'll try to point out as many analogies and motivations from physics as I can see.

  1. If you consider a mechanical system, a term like $\frac{1}{2}k {(x-\mu)^2}$ can occur as measure of elastic energy: For instance when you have a linear spring with spring constant $k$, which can only move in x-direction and is in equilibrium when $X=\mu$. We choose $\sigma$ such that $k=\frac{1}{\sigma^2}$. Larger values of $k$ represent stiffer springs.

  2. Remember that the gibbs measure comes from statistical mechanics, describing an ensemble of particles and where $\beta$ is the inverse temperature: The smaller the value of $\beta$, the faster is the average speed of the particles, i.e. the more likely are states which are "far away" from the equilibrium (in the sense of states with a high energy function $E(x)$). In this way, also $\beta$ represents the "stiffness" of the ensemble of particles.

  3. The terms $\frac{1}{\sqrt{2 \pi \sigma^2}}$ and $\frac{1}{Z(\beta)}$ play a very similar role: They are normalization constants which ensure that the resulting functions are probability densities. These constants naturally depend on $\beta$ and $\sigma^2$.

  4. In conclusion, we can reasonably define $E(X=x)=\frac{1}{2} {(x-\mu)^2}$ as an energy function and set $\beta = \frac{1}{\sigma^2}$ as inverse temperature or measure of stiffness, where smaller values of $\beta$ make high-energy states more likely.

Unfortunately, I cannot provide a meaningful physical interpretation of this particular gibbs ensemble consisting of unidirectional springs.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.