[tylerhou]

Referential transparency, as the original post says. For example, in a functional language you're always allowed to transform

  a = f()
  b = f()
  c = g(a, b)

into

  a = f()
  c = g(a, a)

because f() is guaranteed not to have side effects. (If it does return side effects through the IO monad, instead of actually performing the side effects, g receives two data structures that encode what IO operations to perform.) In an imperative/procedural language, f() may have side effects, and if it does the two programs are not equivalent.

This may (not) make programs easier to reason about.