[BenoitP]

Loops have an implicit ordering, and thus are inherently non parallelizable at face value. map can enable easy parallelization, wether with SIMD or threads or GPU warps or Spark nodes. It's just a construct that specifies less about execution, and thus leaves more room for the compiler/runtime.

This construct is natural with functional programming's abstraction: the function. You encapsulate basic behavior in functions, and then shape the data flow with map(fn)/reduce()/flatmap(). If you really need an index, you can still have them with number ranges, zipWithIndex, etc. If you do use them, you are explicitly reintroducing ordering into your program and lose the parallelism.

This is just a finer level of detail wrt accidental vs essential complexity. IMHO it is always better to specify less about your program, and stick to what you must encode and only that.