home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: talk.origins
- Path: sparky!uunet!zaphod.mps.ohio-state.edu!sdd.hp.com!ux1.cso.uiuc.edu!news.iastate.edu!IASTATE.EDU!kv07
- From: kv07@IASTATE.EDU (Warren Vonroeschlaub)
- Subject: Re: Definition/Information
- Message-ID: <1992Nov22.135134@IASTATE.EDU>
- Sender: news@news.iastate.edu (USENET News System)
- Reply-To: kv07@IASTATE.EDU (Warren Vonroeschlaub)
- Organization: Ministry of Silly Walks
- References: <iHsFuB1w165w@kalki33>
- Date: Sun, 22 Nov 1992 19:51:34 GMT
- Lines: 102
-
- Sorry for the delay on this, but somebody took out all 30+ books on
- introductory information theory from the school library (they really should have
- a withdrawl limit). I had to use one of the advanced books, which is a little
- harder to read.
-
- In article <iHsFuB1w165w@kalki33>, kalki33!system@lakes.trenton.sc.us writes:
- > DEFINITIONS
- >
- > Let S be a system of events E[1],...,E[n] such that
- >
- > 1) P(E[k]) = p[k] (the probability associated with event E[k])
- >
- > 2) 0<=p[k]<=1
- >
- > n
- > 3) SUM p[k] = 1
- > k=1
- >
- > The self-information of the event E[k] is defined as
- >
- > I(E[k]) = -log p[k] (1.1)
- >
- >
- > The entropy of S is defined as
- >
- > n
- > H(S) = - SUM (p[k] log p[k]) (1.2)
- > k=1
-
- Okay, there are a couple of things I want to make clear.
-
- First (in prediction to where this is going. I hope I am wrong or this will be
- short).
-
- In "Relative Information: Theories and Applications" by G. Jumarie, Q360.J84
- 1990 p6, section 1.3.4 "Information and Thermodynamics"
-
- "To the best of our [researchers] knowledge, there is no counterpart of this
- theory [the second law of thermodynamics] in information theory."
-
- In other words, this definition of entropy does not entail that entropy must
- always increase.
-
- Second.
-
- This is not the only information theory of entropy. This is known as the
- Shannon entropy. The Renyi entropy is also good, especially since it can be
- used to determine the relative information content. this is very important in
- modeling events in a specific language (like the physical laws). The Renyi
- entropy is
- n
- Hc(alpha)=(1-c)^-1 ln SUM p[k]^c
- k=1
- where c is the efficiency of detection of the observer. An interesting point is
- that if c>1 then this is a measure of information, and if c<1 this is a measure
- of uncertainty. (I assume to determine the probability of evolution you are
- going to try to calculate the Shannon uncertainty).
-
- There is also the cross-entropy, Hartley, Kulmogorov (for fractals), and
- trajectory entropies. And that is just from one book.
-
- Third.
-
- It might be useful to point out how this is used in information theory. A few
- examples are in order:
-
- Streams. If we have a length of symbols, and we wish to find the entropy of
- the stream, we can use the Shannon entropy as such: A completely random stream
- would be split evenly among the symbols. The Shannon entropy of binary strings
- would then be 1. A stream consisting solely of 1s or 0s would have an entropy
- of 0 (try it). But, if our stream was 0111100011101, the shannon entropy would
- be .9612 (-5/13 *log(5/13) -8/13 *log(8/13), all logarithms base 2). If we do
- this for alternations between two states however, we get 2.4142. So the choice
- of symbols is important. The choice of logarithms isn't. CS people prefer log
- base 2, but ln is poular too.
-
- Descision processes. Let us say we wanted to find the most significant factor
- for car color. Well, then, for each car color we could do the following:
- Separate all the cars according to a factor (say manufacturer), then look at
- what percentage of the cars in each catagory are of the particular color, and
- make that our p[k] for the Shannon entropy (ie p[mazda] for red would be the
- proportion of mazdas that are red). This gives us a Shannon entropy for each
- color car as divided by manufacturer. Take each Shannon entropy for the color,
- and multiply it by the fraction of cars overall that are that color. Add this
- all up to get the entropy of car color by manufacturer.
- If we repeat this for every possible catagory, we will have entropies
- calculated for manufacturer, cost, area, use, etc. The factor that has the
- lowest entropy is the most important factor in car color. Neat huh?
- This has an incredible power as a completely objective way to determine what
- is important in any situation.
-
- Hopefully these examples will be enough to make it clear when Shannon entropy
- is being misused.
-
- | __L__
- -|- ___ Warren Kurt vonRoeschlaub
- | | o | kv07@iastate.edu
- |/ `---' Iowa State University
- /| ___ Math Department
- | |___| 400 Carver Hall
- | |___| Ames, IA 50011
- J _____
-