Principle of maximum entropy

The principle of maximum entropy is a method for analyzing the available information in order to determine a unique epistemic probability distribution. Claude E. Shannon, the originator of information theory, defined a measure of uncertainty for a probability distribution (H(p) = - Σ p_i log p_i) which he called information entropy. In his work, information entropy was determined from (i.e. was a function of) a given probability distribution. The principle of maximum entropy tells us that the converse is also possible: a probability distribution can be determined using the information entropy concept. It states the probability distribution that uniquely represents or encodes our state of information is the one that maximizes the uncertainty measure H(p) while remaining consistent with our information.

The principle of maximum entropy is only useful when all of our information is of a class called testable information. A piece of information is testable if we can determine whether or not a given distribution is consistent with it. For example, the statements

"The expectation of the variable x is 2.87"

and

"p₂+p₃ > 0.9"

are statements of testable information.

Given testable information, the maximum entropy procedure consists of seeking the probability distribution which maximizes information entropy, subject to the constraints of the information. This constrained optimization problem is typically solved using the method of Lagrange multipliers.

Entropy maximization with no testable information takes place under a single constraint: the sum of the probabilities must be one. Under this constraint, the maximum entropy probability distribution is the uniform distribution,

<math>p_i=\frac{1}{n}\ {\rm for\ all}\ i\in\{\,1,\dots,n\,\}.</math>

The principle of maximum entropy can thus be seen as a generalization of the classical Principle of indifference[?], also known as the principle of insufficient reason.

The foregoing discussion applies only to discrete distributions. For continuous distributions, E.T. Jaynes (1963, 1968) finds that the limiting form of the entropy expression as the distribution approaches a continuous distribution is

<math>H_c=-\int p(x)\log\frac{p(x)}{m(x)}\,dx</math>

where m(x), which Jaynes called the "invariant measure", is proportional to the limiting density of discrete points. This invariant measure is not determined by the principle of maximum entropy, and must be determined by some other logical method, such as the Principle of transformation groups[?] or Marginalization theory[?]. Given such an invariant measure, entropy maximization again proceeds by the method of Lagrange multipliers.

Table of contents 1 General Solution of the Maximum Entropy Distribution (Discrete Case) 2 Justifications for the principle of maximum entropy 2.1 Information entropy as a measure of 'uninformative-ness' 2.2 Information entropy as a meaure of the probability of random distribution assignments 3 References

General Solution of the Maximum Entropy Distribution (Discrete Case)

We have some testable information I about a quantity x ∈ {x₁, x₂,..., x_n}. We express this information as m constraints on the expectations of the functions f_k, i.e. we require our epistemic probability distribution to satisfy

<math>\sum_{k=1}^n \Pr(x_i|I)f_k(x_i) = F_k \qquad k = 1, \cdots,m</math>

Furthermore, the probabilities must sum to one, giving the constraint

<math>\sum_{k=1}^n \Pr(x_i|I) = 1</math>

The probability distribution with maximum information entropy subject to these constraints is

<math>\Pr(x_i|I) = \frac{1}{Z(\lambda_1,\cdots, \lambda_m)} \exp\left[\lambda_1 f_1(x_i) + \cdots + \lambda_m f_m(x_i)\right]</math>

with normalization constant

<math> Z(\lambda_1,\cdots, \lambda_m) = \sum_{i=1}^n \exp\left[\lambda_1 f_1(x_i) + \cdots + \lambda_m f_m(x_i)\right]</math>

(Interestingly, the Pitman-Koopman theorem[?] states that the necessary and sufficient condition for a sampling distribution to admit sufficient statistics is that it have the general form of a maximum entropy distribution.)

The λ_k parameters are Lagrange multipliers whose particular values are determined by the constraints according to

<math>F_k = \frac{\partial}{\partial \lambda_k} \log Z(\lambda_1,\cdots, \lambda_m)</math>

These m simultaneous equations do not generally possess a closed form solution, and are usually solved by numerical methods.

Justifications for the principle of maximum entropy

Proponents of the principle of maximum entropy justify its use in assigning epistemic probabilities in several ways, including the following two arguments. These arguments take the use of epistemic probability as given, and thus have no force if the concept of epistemic probability is itself under question.

Information entropy as a measure of 'uninformative-ness'

Consider a discrete epistemic probability distribution among m mutually exclusive propositions. The most informative distribution would occur when one of the propositions was known to be true. In that case, the information entropy would be equal to zero. The least informative distribution would occur when there is no reason to favor any one of the propositions over the others. In that case, the only reasonable probability distribution would be uniform, and then the information entropy would be equal to its maximum possible value, log m. The information entropy can therefore be seen as a numerical measure which describes how uninformative a particular probability distribution is from zero (completely informative) to log m (completely uninformative).

By choosing to use the distribution with the maximum entropy allowed by our information, the argument goes, we are choosing the most uninformative distribution possible. To choose a distribution with lower entropy would be to assume information we do not possess; to choose one with a higher entropy would violate the constraints of the information we do possess. Thus the maximum entropy distribution is the only reasonable distribution.

Information entropy as a meaure of the probability of random distribution assignments

The following argument is essentially the same mathematical argument used in the derivation of the partition function in statistical mechanics, although the conceptual emphasis is quite different. It has the advantage of being strictly combinatorial in nature, making no reference to information entropy as a measure of 'uncertainty', 'uninformative-ness', or any other imprecisely defined concept. The information entropy function is not assumed a priori, but rather is found in the course of the argument.

Suppose an individual wishes to make an epistemic probability assignment among m mutually exclusive propositions. She has some testable information, but is not sure how to go about including this information in her probability assessment. She therefore conceives of the following random experiment. She will distribute N quanta of epistemic probability (each worth 1/N) at random among the m possibilities. (One might imagine that she will throw N balls into m buckets while blindfolded. In order to be as fair as possible, each throw is to be independent of any other, and every bucket is to be the same size.) Once the experiment is done, she will check if the probability assignment thus obtained is consistent with her information. If not, she will reject it and try again. Otherwise, her assessment will be

<math>p_i = \frac{n_i}{N}</math>

where n_i is the number of quanta that were assigned to the i^th proposition.

Now, in order to reduce the 'graininess' of the epistemic probability assignment, it will be necessary to use quite a large number of quanta of epistemic probability. Rather than actually carry out, and possibly have to repeat, the rather long random experiment, our protagonist decides to simply calculate and use the most probable result. The probability of any particular result is the multinomial distribution,

<math>Pr(\mathbf{p}) = W \cdot m^{-N}</math>

where

<math>W = \frac{N!}{n_1 !n_2 !...n_m!}</math>

is sometimes known as the multiplicity of the outcome.

The most probable result is the one which maximizes the multiplicity W. Rather than maximizing W directly, our protagonist could equivalently maximize any monotonic increasing function of W. She decides to maximize

<math>\begin{matrix}\frac{1}{N}\log W

&=& \frac{1}{N}\log \frac{N!}{n_1 !n_2 !...n_m!}\qquad\qquad\qquad\qquad\qquad \\ \\ \ &=& \frac{1}{N}\log \frac{N!}{Np_1 !Np_2 !...Np_m!} \qquad\qquad\qquad\qquad\\ \\ \ &=& \frac{1}{N}\left( \log N! - \sum_{i=1}^m \log Np_i! \right) \qquad\qquad\end{matrix}</math>

At this point, in order simplify the expression, our protagonist takes the limit as N → ∞, i.e. as the epistemic probability levels go from grainy discrete values to smooth continuous values. Using Stirling's approximation, she finds

<math>\begin{matrix}\lim_{N \to \infty}\left(\frac{1}{N}\log W\right)

&=& \frac{1}{N}\left( N\log N - \sum_{i=1}^m Np_i\log Np_i \right)\qquad\qquad\qquad\qquad \\ \\ \ &=& \log N - \sum_{i=1}^m p_i\log Np_i \qquad\qquad\qquad\qquad\qquad\qquad \\ \\ \ &=& \log N - \log N \sum_{i=1}^m p_i - \sum_{i=1}^m p_i\log p_i \qquad\qquad\qquad \\ \\ \ &=& \left(1 - \sum_{i=1}^m p_i \right)\log N - \sum_{i=1}^m p_i\log p_i \qquad\qquad\qquad \\ \\ \ &=& - \sum_{i=1}^m p_i\log p_i \qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad \\ \\ \ &=& H(\mathbf{p}) \qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad \end{matrix}</math>

All that remains for our protagonist to do is to maximize entropy under the constraints of her testable information. She has found that the maximum entropy distribution is the most probable of all "fair" random epistemic distributions, in the limit as the probability levels go from discrete to continuous.

References

Jaynes, E. T., 1963, `Information Theory and Statistical Mechanics,' in Statistical Physics, K. Ford (ed.), Benjamin, New York, p. 181. Available here (http://bayes.wustl.edu/etj/node1).

Jaynes, E. T., 1968, `Prior Probabilities,' IEEE Trans. on Systems Science and Cybernetics, SSC-4, 227. Available here (http://bayes.wustl.edu/etj/node1).