home *** CD-ROM | disk | FTP | other *** search
- Xref: sparky sci.math:15110 sci.math.stat:2364
- Newsgroups: sci.math,sci.math.stat
- Path: sparky!uunet!mcsun!sun4nl!dutrun!donau!dutetvd.tudelft.nl!huub
- From: huub@dutetvd.tudelft.nl (Huub van Roosmalen)
- Subject: Decomposing Markov Models
- Message-ID: <1992Nov17.104641.29410@donau.et.tudelft.nl>
- Sender: news@donau.et.tudelft.nl (UseNet News System)
- Nntp-Posting-Host: dutetvd.et.tudelft.nl
- Organization: Delft University of Technology
- Date: Tue, 17 Nov 1992 10:46:41 GMT
- Lines: 90
-
- Here is a problem ceoncerning Markov Models a friend of mine encountered
- while doing her graduation thesis. Consider a finite state continuous
- time inhomogenous Markov Model with states numbered from 0 to n. Each
- state can be reached in one step from another state. Let the probability
- that the model is in state x be Px. The model represents the total
- activity of a number of videosources on a transmission channel.
-
- In order to simulate this model in hardware, it is highly desirable
- to decompose the original model into n smaller models with only two
- states and zero-state probability Sx. If the original model is in
- state 0, then this implies that all smaller models are in state 0,
- therefore S0*S1*..*Sn = P0. Similarly, if the original model is in
- state n, then this implies that (1-S0)*(1-S1)*...*(1-Sn) = Pn. In
- general, if the original model is in state x, the the probability
- that this occurs equals the probability that x out of n small models
- are in state 1. Thus, a system with states 0 through 7 is to be
- decomposed into 7 systems with two states.
-
- 0-1
- 0-1
- 0-1
- 0-1-2-3-4-5-6-7 -> 0-1
- 0-1
- 0-1
- 0-1
-
-
- To perform the decomposition, we take a z-transform of the original
- model's state probability density. The roots of the polynomial of the
- z-transform correspond to the state probabilities of the decomposed
- model. The reasoning behind this is that if we have the decomposed
- model, then the distribution of the sum of the states of the small
- models equals the convolution of their state distributions. In the
- z-domain this is equivalent to multiplication, so if we factorize the
- z transform of the original model, then the factors should each
- represent the z-transform of a small model.
-
- A typical z-transform would be:
-
- 3 + 31 z + 33 z^2 + 28 z^3 + 22 z^4 + 23 z^5 + 6 z^6 + z^7
- ---------------------------------------------------------- (1)
- 147
-
- The relationship between the roots Zk and the probabilities of the
- smaller models being in state 0 is:
-
- -Zk
- Sk = -------- (2)
- 1 - Zk
-
-
- The trouble begins when we calculate the roots: some of them turn out
- to be complex valued. Inserting these complex values into Equation 2
- leads to complex probabilities!
-
- One observation we could make is that since the coefficients of the
- polynomial are real, the complex zeros must be complex conjugated
- values. We could try to settle for less by trying to decompose the
- original model into a set of models with two states for real zeros
- and three states for complex zeros. In this case the example with
- 7 states in the original model could, for one real zero and three
- complex pairs, be decomposed as:
-
- 0-1
- 0-1-2
- 0-1-2-3-4-5-6-7 -> 0-1-2
- 0-1-2
-
- The results are slightly better, but we still end up with negative
- values for the probability that a small three-state model is in state 1
- and the probabilities for states 0 and 2 of the small models are real
- and positive and add up to 1.
-
- I have several questions on this:
-
- - Is it possible at all to perform a decomposition of the model in
- this way?
-
- - If so, How should I interpret (transform?) the complex or negative
- values resulting from this model? Taking the modulus or just the
- real or imaginary parts of the results does not seem to work
-
- - Are there any references available to literature treating complex
- or negative probabilities? All pointers are welcome
-
- Huub van Roosmalen
- huub@dutetvd.et.tudelft.nl
- Faculty of Electrical Engineering
- Delft University Of Technology
- The Netherlands
-