Vectorization is the practice of expressing a computation as operations on whole arrays rather than as explicit element-by-element loops in the host language. In Python, writing a loop that adds two million-element lists is slow because every iteration pays the cost of the interpreter: bytecode dispatch, type checks, and object boxing. Vectorized code says “add these two arrays” in a single expression, and the actual loop runs inside a compiled routine where none of that overhead exists.
NumPy is the canonical example, and its glossary is explicit about the mechanism: “NumPy hands off array processing to C, where looping and computation are much faster than in Python. To exploit this, programmers using NumPy eliminate Python loops in favor of array-to-array operations.” The term, the glossary notes, refers both to the C offloading itself and to the practice of structuring code to take advantage of it. The Python interpreter is involved only once - to dispatch the whole-array operation - and then steps out of the way.
The engine behind this in NumPy is the universal function, or ufunc. The documentation defines a ufunc as a “vectorized wrapper for a function that takes a fixed number of specific inputs and produces a fixed number of specific outputs,” operating on arrays “in an element-by-element fashion.” When you write a + b on two arrays, NumPy calls the add ufunc, whose inner loop is compiled C that walks the arrays’ memory directly. Because the data is a contiguous block of one dtype, that loop can also exploit the processor’s SIMD instructions, performing the same operation on several elements at once.
Vectorization is therefore as much a discipline as a feature. The programmer’s job becomes finding the array-shaped expression of a problem - a sum, a product, a comparison applied across an entire axis - rather than spelling out the iteration. This often makes the code shorter and clearer as well as faster, since intent is stated at the level of arrays. It is a direct descendant of array-oriented thinking that goes back through Fortran’s numerical libraries.
The technique scales far beyond NumPy. Every deep-learning framework is built on the same idea: operations are defined over whole tensors so that the heavy lifting happens in optimized kernels on the CPU or GPU, with the high-level language acting only as the conductor. Vectorization, broadcasting, and the tensor data structure together are what let interpreted Python drive numerical workloads at compiled-code speeds.