Mathematics Stack Exchange is a question and answer site for people studying math at any level and professionals in related fields. Join them; it only takes a minute:

Sign up
Here's how it works:
  1. Anybody can ask a question
  2. Anybody can answer
  3. The best answers are voted up and rise to the top

Formula is here: $$ p(i)=\frac{e^\frac{f(i)}{T}}{\displaystyle \sum_j e^\frac{f(j)}{T}} $$


1) Each $p(i)$ is a number between $0$ and $1$, no matter what the fitness is (positive or negative). This scheme does not require that fitness has to be positive.

2) The sum of all the $p(i)$'s is $1$, i.e. this is a probability distribution.

3) No matter what $T$ is:

  • If two items have same fitness, they have same probability of being picked.
  • If all fitnesses are the same, we pick random item.

4) No matter what the fitnesses are:

  • As $T\to\infty$ we tend to pick random item.
  • As $T\to0$ we tend to pick only the best item. That is, its probability is $1$, the probability of all others is $0$. If there are $m$ joint best items, we pick them with probability $1/m$, and all others with probability $0$.
share|cite|improve this question
For (2), ... if the denominator is summing up $n$ or so terms, and then you add $n$ of the $p(i)$, well... – J. M. Aug 14 '11 at 17:40
For (1): Why would the denominator be always greater than or equal to the numerator? – J. M. Aug 14 '11 at 17:41 really could have started with defining the $p$, $f$, and $T$ in the formula you gave, though... – J. M. Aug 14 '11 at 17:42

Below is a probably excessively detailed answer. The first few items are plain algebra. The $4$th question needs some knowledge about the behaviour of the exponential function. If you need more detail there, please leave a comment.

To make typing easier, let $$a_i=e^{\frac{f(i)}{T}}.$$ (Actually, it is not just to make typing easier. Simpler looking expressions are easier to grasp.) Let $n$ be the number of terms in your sum. Note that $$p(i)=\frac{a_i}{a_1+a_2+\cdots+a_n}.$$

(1): For any $x$, $e^x \gt 0$. so all of the $a_i$ are positive. If $n=1$ (one term only!), then $p(1)=a_1/a_1=1$. If $n>1$, then since the $a_i$ are positive, $a_i \lt \sum a_i$, so $p(i)=a_i/(a_1+\cdots +a_n)\lt 1$.

(2):The sum $p(1)+p(2)+\cdots +p(n)$ is equal to $$\frac{a_1}{a_1+a_2+\cdots +a_n}+\frac{a_1}{a_1+a_2+\cdots +a_n}+\cdots+\frac{a_n}{a_1+a_2+\cdots +a_n}.$$ Thus $$p(1)+p(2)+\cdots +p(n)=\frac{a_1+a_2+\cdots+a_n}{a_1+a_2+\cdots +a_n}=1.$$ More compactly, $$\sum p(i)=\sum \frac{a_i}{\sum a_i}=\frac{\sum a_i}{\sum a_i}=1.$$

(3): If $f(i)=f(j)$ (same fitness) for some particular $i$ and $j$ then $a_i=a_j$, so $p(i)=p(j)$. The fact that $p(i)=p(j)$ means that the probability that $i$ is picked is the same as the probability that $j$ is picked.

If $f(i)=f(j)$ for all $i$, $j$, that means that $p(i)=p(j)$ for all pairs $(i,j)$, meaning that everyone has the same probability of being picked.

(4): This is the only part that requires that we look in more detail at the definition.

Limit as $T\to \infty$: As $T\to \infty$, $f(i)/T \to 0$, so $e^{f(i)/T}\to 1$. (Recall that $e^0=1$.) So each $a_i$ approaches $1$. This means that each $p(i)$ approaches $1/(1+1+\cdots +1)$. So each $p(i)$ approaches $1/n$. Thus, when $T$ is very large, all items have nearly equal probability of being picked.

Limit as $T\to 0$: First a technical remark. Look at $e^{1/T}$ when $T$ is close to $0$. If $T$ is positive, then $e^{1/T}$ is huge. If $T$ is very large negative, then $e^{1/T}$ is close to $0$. So it matters very much on which side of $0$ we are! I will assume that the problem really means to ask about what happens when $T$ approaches $0$ from the right, that is, through positive values.

Suppose that there is a single best item. For definiteness, assume that item $1$ is the unique best item. Then $f(1) \gt f(i)$ for all $i\gt 1$. The probability of picking item $1$ is $p(1)$, where $$p(1)=\frac{e^{\frac{f(1)}{T}}} {e^{\frac{f(1)}{T}}+e^{\frac{f(2)}{T}}+\cdots+e^{\frac{f(n)}{T}}}.$$ Divide "top" and "bottom" by $e^{\frac{f(1)}{T}}$, or equivalently multiply top and bottom by $e^{-\frac{f(1)}{T}}$. We obtain $$p(1)=\frac{1} {1+ e^{\frac{f(2)-f(1)}{T}}+ e^{\frac{f(3)-f(1)}{T}}\cdots+e^{\frac{f(n)-f(1)}{T}}}.$$

Finally, let $T$ approach $0$ from the right. Look now for example at the term $e^{\frac{f(2)-f(1)}{T}}$ in the denominator, or any of the other terms in the denominator except for the initial $1$. Recall that since item $1$ is the unique most fit one, we have $f(2) \lt f(1)$, and therefore $f(2)-f(1)<0$. So when $T$ is very close to $0$ but positive, $e^{\frac{f(2)-f(1)}{T}}$ is very close to $0$, since we are looking at an $e^y$ where $y$ is large negative.

So as $T$ approaches $0$ from the right, all of the denominator dies except for the first term, and therefore $p(1)$ approaches $1$.

Almost exactly the same idea will work if there are $m$ "fittest" items, say items $1$ to $m$. The difference is that when we divide top and bottom, there will be $m$ $1$'s in the denominator, not just the single $1$ that we had in the unique best item case.

share|cite|improve this answer

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

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