Running totals
Create running totals from a one-dimensional sequence such as transaction amounts.
NumPy · Function
np.cumsum
Computes cumulative sums along an axis of an array. By default, multidimensional input is flattened before accumulation.
How it works
NumPy converts the input to an array and accumulates sums along the selected axis. With no axis, it processes the flattened input; dtype controls the accumulator and result type, while out can receive the result in a supplied array.
Syntax
numpy.cumsum(a, axis=None, dtype=None, out=None)Parameters
| Name | Type | Description |
|---|---|---|
a | array_like | Input array. |
axis | int, optional | Axis along which the cumulative sum is computed. The default, None, computes over the flattened array. |
dtype | dtype, optional | Type of the returned array and the accumulator used for summation. |
out | ndarray, optional | Array in which to place the result; it must have the expected shape and buffer length, with casting applied if needed. |
Returns
An ndarray containing cumulative sums, with the same size as the input; its shape matches the input when axis is specified or the input is one-dimensional. If out is supplied, a reference to out is returned.
Key points
- Produces running totals rather than one aggregate total.
- For an explicitly selected axis, the output retains the input shape.
- dtype affects both the output array and the internal accumulator.
- out lets callers reuse or provide storage for the computed values.
- A multidimensional input keeps its shape only when axis is specified; a one-dimensional input already has that form.
Common mistakes
- Leaving axis unset changes a multidimensional array into a one-dimensional cumulative sequence.
- An out array must have the expected shape and buffer length.
- Integer accumulation can wrap around on overflow instead of raising an exception.
Example: Cumulative sums across rows
import numpy as np
a = np.array([[1, 2, 3], [4, 5, 6]])
print(np.cumsum(a, axis=1))
# [[ 1 3 6]
# [ 4 9 15]]Computes running totals independently across each row.
When to use it
Per-axis accumulation
Generate row-wise or column-wise cumulative measurements in a matrix.
Output-buffer reuse
Write results into preallocated NumPy storage to reduce additional allocation.
Interview questions
QWhat does np.cumsum compute, and what does axis=None mean?
It computes cumulative sums along an axis, where each output element includes the current input element and all preceding elements along that axis. If axis is None, the input is flattened first.
Follow-up: How would you preserve row structure for a 2-D array? Pass axis=1 for running totals within rows, or axis=0 for running totals down columns.
QWhat is the purpose of dtype?
dtype selects the type of both the returned array and the accumulator used during summation. Choosing a suitable dtype can control precision or reduce integer-overflow risk.
Follow-up: What caveat remains for floating-point inputs? Cumulative addition can differ from a separately grouped sum because floating-point arithmetic is not exact.
QWhen would you use out?
out is an optional ndarray that receives the result instead of requiring a newly allocated output. It must match the expected shape and buffer length, and values are cast to its type when necessary.
Follow-up: What should be checked when using out? Ensure its shape is compatible and its dtype can represent the intended result without unwanted casting.
Related
numpy.ndarray.cumsumnumpy.sum