NumPy Broadcasting Rules — ELI5
Imagine you have a recipe that serves four people, but tonight you are cooking for twelve. You do not rewrite the recipe three times — you just multiply every ingredient by three. Your brain “stretches” the recipe to fit the bigger number.
NumPy does exactly the same trick. When you ask it to add a small array to a big array, it stretches the small one so the shapes match, then does the math. That stretching is called broadcasting.
Say you have a table of test scores for thirty students across five subjects. You want to add five bonus points — one different bonus per subject. Instead of building a full thirty-by-five bonus table, you hand NumPy a tiny list of five numbers. NumPy copies that list down every row automatically.
The rule is simple: line up the shapes from the right. If a size is one, stretch it. If sizes already match, great. If neither of those is true, NumPy refuses and gives you an error — it will not guess what you meant.
People often think broadcasting copies data in memory. It does not. NumPy is clever enough to pretend the data is repeated without actually using extra space. That is why broadcasting is fast, not just convenient.
The one thing to remember: Broadcasting stretches smaller arrays to match bigger ones — no loops, no extra memory, no hassle.
See Also
- Python Bokeh Get an intuitive feel for Bokeh so Python behavior stops feeling unpredictable.
- Python Numpy Advanced Indexing How to cherry-pick exactly the data you want from a NumPy array using lists, masks, and fancy tricks.
- Python Numpy Einsum One tiny function that replaces dozens of NumPy operations — once you learn its shorthand, array math becomes a breeze.
- Python Numpy Fft Spectral How NumPy breaks apart a signal into its hidden frequencies — like separating a chord into individual notes.
- Python Numpy Memory Views Why NumPy arrays can share the same data without copying it — and how that makes your code fast but occasionally surprising.