>>479 つづき

１）
さて、まず、[HT08b] より
(抜粋)
P91
1. INTRODUCTION.
We often model systems that change over time as functions from the real numbers R (or a subinterval of R) into some set S of states, and it is often our goal to predict the behavior of these systems.
Generally, this requires rules governing their behavior, such as a set of differential equations or the assumption that the system (as a function) is analytic.
With no such assumptions, the system could be an arbitrary function, and the values of arbitrary functions are notoriously hard to predict.
After all, if someone proposed a strategy for predicting the values of an arbitrary function based on its past values, a reasonable response might be,
“That is impossible. Given any strategy for predicting the values of an arbitrary function, one could just define a function that diagonalizes against it: whatever the strategy predicts, define the function to be something else.”
This argument, however, makes an appeal to induction:
to diagonalize against the proposed strategy at a point t, we must have already defined our function for all s < t in order to determine what the strategy would predict at t.

In fact, the lack of well-orderedness in the reals can be exploited to produce a very counterintuitive result: there is a strategy for predicting the values of an arbitrary function, based on its previous values, that is almost always correct.
Specifically, given the values of a function on an interval (?∞, t), the strategy produces a guess for the values of the function on [t,∞), and at all but countably many t, there is an ε > 0 such that the prediction is valid on [t, t + ε).
Noting that any countable set of reals has measure 0, we can restate this informally: at almost every instant t, the strateg predicts some “ε-glimpse” of the future.

つづく