Vectorized operations and broadcasting

Apply element-wise calculations to whole arrays and use compatible shapes without writing Python loops.

Vectorize element-wise calculations

NumPy arithmetic and many functions operate element by element over an array. This style is called vectorization; it expresses the calculation on the whole array instead of writing a Python loop for every value:

readings = np.array([18.0, 20.5, 23.0])
fahrenheit = readings * 9 / 5 + 32

print(fahrenheit)  # [64.4 68.9 73.4]
print(np.round(fahrenheit))  # [64. 69. 73.]

Universal functions, or ufuncs, include operations such as np.sqrt, np.exp, and np.add. Comparisons are element-wise too:

values = np.array([1.0, 4.0, 9.0])
print(np.sqrt(values))       # [1. 2. 3.]
print(np.isfinite(values))   # [ True  True  True]

Common element-wise functions:

  • np.add, np.subtract, np.multiply, np.divide: arithmetic counterparts of +, -, *, /.
  • np.mod, np.power: counterparts of % and **.
  • np.reciprocal: element-wise reciprocals; use a floating dtype when fractional results are expected.
  • np.square, np.sqrt: square and square-root each value.
  • np.sin, np.cos: trigonometric functions; inputs are in radians.

For numeric arrays, * means element-wise multiplication, not matrix multiplication:

left = np.array([1, 2, 3])
right = np.array([10, 20, 30])
print(left * right)  # [10 40 90]

Let shapes broadcast

Broadcasting lets operations combine shapes when their dimensions are compatible. Compare shapes from the right: dimensions must be equal, or one of them must be 1. A missing leading dimension behaves like 1.

observations = np.array([[10.0, 20.0, 30.0],
                         [12.0, 18.0, 27.0]])
offsets = np.array([1.0, 2.0, 3.0])

adjusted = observations + offsets
print(adjusted.shape)  # (2, 3)
print(adjusted)

The one-dimensional offsets array aligns with the final axis, so each row receives the same column-wise offsets. Add an axis explicitly if you want to combine each row value with every value from another vector:

row_values = np.array([10, 20])
column_values = np.array([1, 2, 3])
all_pairs = row_values[:, np.newaxis] + column_values

print(all_pairs.shape)  # (2, 3)

If two dimensions differ and neither is one, NumPy raises a ValueError. Print both shapes and check their trailing dimensions before trying to reshape data.

Exercise: Convert a (3, 4) table from meters to centimeters with one array expression. Then add a four-element per-feature offset. Try an incompatible shape and explain the error.

Sources: Universal functions, Mathematical functions, Broadcasting