Practical Constraints on Reason

Background
• From the Church-Turing thesis, we now have a clear statement of what an effective procedure is.
• Many questions follow:
• Are there things that we might like to know but which a Turing machine cannot do?
• Are there any measures on effective procedures which we may want or need to add to our understanding of computation?
• The answer to both of these questions is yes.
• We have already seen one problem no Turing machine can solve that we want to solve: an effective way to decide whether any effective procedure will give an answer (i.e., whether a machine will halt).
• These kinds of limitations (Godel's theorems, the halting problem) are in principle limitations on reason.
• There are also practical limits on reason. These include:
1. How long it takes to do something (time complexity).
2. How much memory it takes to do something (space complexity).
3. Sensitivity to initial conditions (chaos).
1. Time Complexity
• Call an effective procedure an "algorithm."
• A Turing machine takes discrete steps. Assuming that for any real implementation, these take time, a measure of the number of steps is a measure of time complexity.
• Time complexity is measured as a function of the input. An input of n items may require n, 2n, n2, or even n! steps. The more steps it requires, the more time.
• Some effective procedures require so much time for certain tasks that they are practically impossible to solve (although useful approximations may be possible).
• Complexity is measured with "O-notation." This is to say some task is bounded by some function plus some constant. To be O(2n), for example, means that for n items there will be as many or fewer than 2n + K steps for some constant K.
• An example of a very time consuming but useful problem is the travelling salesman problem. This problem is "NP- complete," meaning that it is in a special category of problems each of which is equivalently very difficult and apparently cannot be done in polynomial time (for input n, it appears to be worse than any solution of nk for some constant k).
• A fun puzzle is the Tower of Hanoi: disks of various sizes are stacked by order of size on one of three rods in a line; you must move them one at a time from the left rod where they are stacked to the rightmost one. BUT, you cannot stack a disk on top of a smaller disk. This problem is of the order of 2n steps for n disks. In class, I displayed a variant where you could also only move the disks one post at a time; Justin noted that this could be described recursively where f(0) = 0, and for each n there are 3f(n) + 2 steps.
2. Space complexity
• For an input of n items, an algorithm will also require working space (for our Turing machines, this is tape).
• We measure space complexity as the amount of memory (tape squares) that the algorithm will take, as a function of an input of n items.
• Like time complexity, some problems require impractical amounts of memory, and cannot therefore ever really be solved.
3. Chaos
• By "chaos," we mean systems that are very sensitive to initial conditions. Such a system might be one where a very slight change now results in a very big change later.
• The problem with such systems is that if they are very sensitive, then errors in measurement can in time result in predictions being very inaccurate. Suppose for example that some aspect of the weather is choatic. You measure the temperature and wind velocities and etc., and feed this into a predictive model. But if the model included chaotic equations, then even if it were accurate, some tiny error in measurements might result in the prediction being very far off from what actually results.
• A simple example is the logistic equation, an iterative or recursive equation: xn = K*xn-1*(1 - xn-1 ). n refers to the step in the iteration. x should be between 0 and 1. When K is around 3.6+, this equation is sensitive to initial conditions over time.