1) If X is at its mean then Y is at its mean. Like you say when X goes up from its mean then Y will go down from its mean.

2) 1000X is not normally distributed. If X's population was {1,0,0,1,0,1} then 1000X's population would be {1000,0,0,1000,0,1000} it still only has two values.

3) We correct for things such as standard error for a finite population with a "finite correction factor" (Where N is population size and n is sample size) if N increases then you do need to increase n to maintain the same accuracy.

You can read more about this here http://courses.wcupa.edu/rbove/Beren...section7_3.pdf

4) Decreasing marginal returns tends asymptotically to zero. You cannot use a linear relationship to model that.