The Broadcasting Rules

Broadcasting is not magic and not guesswork — NumPy follows a short, strict algorithm. Once you can run this algorithm in your head, you can predict every broadcasting result (and every error) before you run the code.

Definition

Two shapes are broadcast-compatible when, comparing their dimensions from right to left, every pair of dimensions is either equal or one of them is 1. Missing dimensions on the left are treated as 1. The result shape takes the larger size in every position.

• Right to left — NumPy lines shapes up at their last (trailing) dimension, not the first.
• Equal or 1 — the only two ways a pair can be compatible. 5 vs 5 is fine, 1 vs 14 is fine, 5 vs 14 is an error.
• Missing dimensions = 1 — a (14,) array next to a (5, 14) array is treated as (1, 14).
• Larger size wins — the size-1 side is stretched to the other size.

The rules step by step

Step 1  Write the two shapes right-aligned, one under the other.
Step 2  If one shape is shorter, pad it on the LEFT with 1s.
Step 3  Compare each column (right to left):
          equal          -> OK, keep that size
          one of them 1  -> OK, use the bigger size (stretch the 1)
          anything else  -> ValueError: cannot broadcast
Step 4  The result shape is the bigger size in every column.

Worked example — (5, 14) with (14,):

A:  5  14
B:     14      -> pad left ->  1  14
               compare:        5 vs 1  -> OK (stretch to 5)
                              14 vs 14 -> OK (equal)
Result: (5, 14)

Worked example — (5, 14) with (5,) (fails):

A:  5  14
B:      5      -> pad left ->  1   5
               compare:       14 vs 5  -> neither equal nor 1 -> ERROR

Worked example — 3D (2, 1, 3) with (1, 4, 3):

A:  2  1  3
B:  1  4  3
    2  4  3   <- 3 vs 3 equal, 1 vs 4 stretch, 2 vs 1 stretch
Result: (2, 4, 3)

Note: Both arrays can be stretched at the same time. In (3, 1) + (1, 3) the first stretches across columns and the second down rows, giving (3, 3).

Syntax — check shapes without doing the math

np.broadcast_shapes(shape_a, shape_b, ...)   # returns result shape tuple

Example — let NumPy apply the rules

import numpy as np

print(np.broadcast_shapes((5, 14), (14,)))
print(np.broadcast_shapes((5, 14), (5, 1)))
print(np.broadcast_shapes((3, 1), (1, 3)))
print(np.broadcast_shapes((2, 1, 3), (1, 4, 3)))
print(np.broadcast_shapes((2, 3), ()))

(5, 14)
(5, 14)
(3, 3)
(2, 4, 3)
(2, 3)

np.broadcast_shapes((5, 14), (5,))

ValueError: shape mismatch: objects cannot be broadcast to a single shape.  Mismatch is between arg 0 with shape (5, 14) and arg 1 with shape (5,).

How it works: np.broadcast_shapes runs exactly the four steps above, without creating any arrays. It is the fastest way to answer "will this broadcast?" during debugging.

Compatibility table

Every row below was checked with np.broadcast_shapes:

Tip: Notice the row (4, 1) with (4,) → (4, 4). This is a famous silent surprise: no error, but the result is a 4×4 grid, not 4 values. Always check .shape of the result.

Common mistakes

• ❌ Aligning shapes from the left (5 with 5 in (5, 14) vs (5,)). ✅ Always align from the right.
• ❌ Thinking "any small number stretches". ✅ Only a 1 (or a missing dimension) stretches. A 2 never stretches to 4.
• ❌ Assuming the result shape equals the first operand's shape. ✅ The result takes the max of each column — it can be bigger than both inputs, e.g. (3, 1) + (1, 3) = (3, 3).

Interview tip

Q: Will (8, 1, 6, 1) broadcast with (7, 1, 5)? What is the result? A: Pad B to (1, 7, 1, 5). Compare right to left: 1 vs 5 → 5, 6 vs 1 → 6, 1 vs 7 → 7, 8 vs 1 → 8. Result (8, 7, 6, 5).