## Induction

### August 28, 2010

Let the world of sense data available to an observer be an infinite sequence of bits. At time 0 the observer wishes to verify or refute a hypothesis about the sequence. The hypothesis is a prediction of the entire string from time 0 on. The question of induction is whether the observer can, by verifying that the pattern holds until a certain time n, make it more likely that the pattern holds beyond time n, and possibly holds for ever, without making any a-priori assumptions whatsoever about the sequence.

Clearly, if n is any fixed number then the answer is negative. For any n, and any hypothesis, there is a sequence (actually, an infinite number of sequences) for which the hypothesis holds for bits 1 through n, and fails on bit n + 1. Thus, the observer is led to adopt the hypothesis only to have it fail immediately upon adoption.

If, however, the observer can generate random bits, then some induction is possible. In this case the observer can generate a random number n which is distributed over the natural numbers and which has a probability pk of being equal to k. If the sequence fits the hypothesis to bit m, then the observer will only adopt the hypothesis in the case n < m and, if so, will make at least one successful inductive prediction unless n = m – 1. Thus, the probability of making an immediate error is pm. Given the distribution, pk, k = 1, 2, …, the highest chance for immediate failure occurs when pm maximizes pk over k = 1, 2, …. Thus, the minimax strategy for the probability of immediate failure is to minimize the maximum pk. Under the constraint that the expected value for n should be no more than a certain value, b, the minimax strategy for the observer is to sample n uniformly over the range 1, 2, …, 2b. The probability of adoption and immediate failure in this case is no more than 1 / 2b, regardless of what the sequence is.

Similarly, the chance of hypothesis adoption and failure within d predictions is d / 2b. Being able to predict with any confidence that a certain hypothesis will hold forever is clearly impossible no matter how many trials one is allowed to make before adopting the hypothesis.

Similar analysis holds, mutatis mutandis, if the sense-data is assumed to be a observed continuously in time. Observing the hypothesis holding for a certain random period of time enables the observer to predict with some confidence that the hypothesis would continue to hold for some comparable length of time beyond that observation period.

## Danilo Zolo on eklogecratic tools

### August 9, 2010

Danilo Zolo’s book Democracy and Complexity (1992) is a critique of the “classical” and the “neo-classical” doctrines of democracy. The classical theory espouses the idea that in a democracy the entire population of a country, en masse, can – in some way, usually assumed to rely heavily on elections – determine public policy. The neo-classical doctrine, sometimes referred to as the Schumpeterian theory, asserts that the people do not, and indeed cannot, determine public policy. According to this theory, democracy is merely the situation in which people are able to occasionally select, through “free” elections, which elite group is to set public policy.

Zolo agrees with the neo-classical argument against the classical theory, namely, that a large population does not, and cannot, determine policy. He is, however, critical of the neo-classical theory as well. He argues that for various reasons (resulting from “complexity”) elections are not only not a tool for setting policy, they cannot even be seen as reflecting popular choice of an alternative among elites.1
Read the rest of this entry »