# "back substitution" method for solving linear system

I'm using numpy to write the "back substitution" method for solving linear system where "A" is a nonsingular upper triangular matrix.

import numpy as np

def upperTriSol(A, b):
n = np.size(b)
x = np.zeros_like(b)

x[-1] = 1. / A[-1, -1] * b[-1]
for i in xrange(n-2, -1, -1):
x[i] = 1. / A[i, i] * (b[i] - np.sum(A[i, i+1:] * x[i+1:]))

return x


I know that "for loops" are slow, so I wonder if there is any way to avoid them in this case? If not, is there a "correct" way to write efficient "for loops" in numpy?

• What was wrong with numpy.linalg.solve? Nov 6 '15 at 19:30
• @GarethRees Nothing wrong with it. It's just some college stuff we need to write by ourselves. It works already, just want to improve a little if possible. Nov 6 '15 at 21:44
• Can you edit the question to explain that, please? Nov 6 '15 at 22:24

If looks like x[i] is a function of x[i+1] and selected items from A and b. Then it could also be seen as forward itertion on a reversed list, xr[i+1] dependent on xr[i].
np.cumsum and related accumulated functions are often useful in problems like this.
There's a set of tri functions that give you indices and values on upper and lower triangular matrices.
Typically 'vectorizing' in numpy requires stepping back a bit from the problem and imagining in terms of operations on a table or 2d array of values.