Create normally distributed arrays
Generate synthetic Gaussian-valued features or noise with a specified tensor shape.
NumPy · Function
np.random.randn
Draw samples from the standard normal distribution. With dimensions, np.random.randn returns an array of that shape; without dimensions, it returns one float.
How it works
Call np.random.randn with zero or more dimension arguments. With no dimensions it returns one Python float; with positive integer-like dimensions it returns an ndarray whose shape follows those arguments, containing samples from a normal distribution with mean 0 and variance 1.
Syntax
numpy.random.randn(d0, d1, ..., dn)Parameters
| Name | Type | Description |
|---|---|---|
d0, d1, ..., dn | int, optional | Dimensions of the returned array; dimensions should be positive integer-like values. If omitted, one scalar is returned. |
Returns
ndarray or float: an array shaped (d0, d1, ..., dn) containing floating-point standard-normal samples, or one float when no parameters are supplied.
Key points
- The function draws one or more values from the standard normal distribution.
- Array elements are random floating-point Gaussian values with mean zero and variance one.
- Supplying dimensions makes the result shape match the positional dimension sequence.
- With no arguments, the result is one scalar float.
- A nonstandard normal can be formed by scaling the samples by sigma and shifting them by mu.
Common mistakes
- Calling randn() produces a scalar float, not a zero-dimensional ndarray.
- Dimension arguments must be positive; use an explicit shape with the Generator API when you need different semantics.
Example: Generate standard-normal samples
import numpy as np
rng = np.random.default_rng(42)
z = rng.standard_normal((2, 3))
print(z.shape) # (2, 3)
sample = np.random.randn()
print(isinstance(sample, float))Creates a reproducible standard-normal array with the recommended generator method, then demonstrates that randn() without dimensions returns one scalar.
When to use it
Draw a scalar
Produce one standard-normal deviate for a stochastic calculation.
Transform a standard normal
Construct N(mu, sigma²) values by scaling and shifting standard-normal output.
Interview questions
QWhat distribution does np.random.randn use?
It samples from the standard normal distribution, so the mean is 0 and the variance is 1. With dimensions, it returns floating-point samples arranged in the requested shape.
Follow-up: How would you sample from N(mu, sigma²)? Multiply the output by sigma and add mu: mu + sigma * np.random.randn(...).
QHow does the return type depend on the arguments?
No arguments return one Python float. Passing d0, d1, and so on returns an ndarray with shape (d0, d1, ..., dn).
Follow-up: What is the modern alternative? Use a Generator instance and call its standard_normal method, passing a tuple or integer shape.
QHow do you request a multidimensional result?
Each dimension is supplied as a separate positional argument, such as np.random.randn(2, 4), producing a 2-by-4 array. The dimensions should be positive integer-like values.
Follow-up: How can the same shape be expressed with the modern API? Call rng.standard_normal((2, 4)) on a Generator instance.
Related
numpy.random.standard_normalnumpy.random.Generator.standard_normal