L-1 Norm in multivariate regression, Basic Principles
			by Jon (6/19/2009)


r = [1]m - d    gives three absolute value functions, one equation in basis
r = [b]m - d    scaled absolute value functions

r = [a b] -d    two equations in basis.

		Claerbout's method  (exact solution in finite number of steps)
			take gradient
			descend via weighted median
			find a new basis equation
			throw out oldest basis equation

		This method is of unknown quality
		compared to linear programming and new methods.

		This method needs revision for problems
		of very high dimensionality.

Next:   REGULARIZATION
0 \approx (1,-1)* m 
		With a big enough epsilon, some of these equations
		will be in the basis, i.e. step functions in the solution.

----------------------------------------------------------------------

	What applications do we want to consider?

	Can we make very simplified synthetics of them?

	Can we test all our algorithms on all our synthetics?