Monday, January 31, 2011

Quantization and rank

Playing around with the the yield curve, we ask, what happen when the economy drops the 'stages of production', the rank.  Quantization error goes up.  Here I use term values chosen from a Fibonacci sequence and find that the accuracy increases as we increase rank.

The economy has dropped rank once, and is likely to drop again, so as we get simpler we get less accurate.  But there is another curve called gains from specialization which decrease as we increase rank. The two are in conflict, we balance precision and risk.

This method uses Shannon (Channel) Theory, looking at sparse sampling compared to perfect sampling.. The Levine paper constructs the balancing equations from scratch.

What is my assumption?  I want the quantization error from each banker along the curve, under the assumption that he has the optimum share of the curve, and is not sharing more than a Gibbs amount of bandwidth with the bankers upstream and down stream.
So, the entropy iseparated as:  Pi* log(Pi), where Pi is the probability of a sample, 1/term, of the banker, his sample rate.
So from Shannon:  C = B log(1+SNR) can be taken for each banker, I have broken up the channel into components of number equal to rank.

My assumptions have no other risk only quantization error and the total SNR is the mean SNR from each banker, so quantization error become 1/(SNR) for each banker.  In the limit, as the number of bankers along the curve grows, the quantization error goes to zero.  Later I will introduce risk as a measure of the over crowding along the yield curve.  There will be a trade off, fewer bankers along the curve yield more quantization error, but less risk of overlap among bankers. Look at the two yield curves presented below.  One has three terms, and one has seven.  
Do bankers split the transaction rates along the curve by Fibonacci? It looks that way, and likely true if bankers are always searching for equilibrium between two ranks taking the shortest number of steps.These yield curves do the same function as bankers, they measure their own error.

No comments: