What makes NumPy broadcasting efficient and when might it waste memory?

Short answer

Broadcasting applies compatible array shapes without explicitly duplicating many input elements, but materializing the expression result and temporary arrays can still consume substantial memory.

Why

NumPy compares dimensions from the right. Dimensions match if equal or if one is 1. A column vector of shape (1000, 1) can combine with a row vector of shape (1, 1000) to produce a million-element result without first copying both inputs into a million-element grid.

The new result itself still needs storage, and chained expressions can create additional temporaries. In bioinformatics, subtracting a sample-level control from a gene-by-sample matrix is often efficient, while accidentally broadcasting across mismatched axes can produce an unintended quadratic array. Check shapes, dtypes and intermediate sizes.

Consider in-place operations only when mutation is safe, and chunk results when output size cannot fit in memory.

Common mistakes

Assuming broadcasting never allocates large arrays or that matching shapes imply the intended biological axis

The interviewer might ask next

What shapes would result from adding arrays of size (5000, 1) and (1, 5000)?