Why does FFT multiplication wrap around?

The finite DFT represents periodic sequences, so its convolution theorem uses circular convolution.

Can it compute linear convolution?

Yes, with sufficient padding and correct output handling.

Is wrap-around always wrong?

No. It is appropriate when periodic boundaries are intended.

AI Glossary / Computer vision and multimodal AI / Intermediate

Circular convolution

Definition

Circular convolution combines two finite sequences while treating their indices as periodic, so values that pass one end wrap around to the other. It is the convolution operation naturally associated with multiplying equal-length discrete Fourier transforms and taking an inverse transform.

What it means

Ordinary linear convolution includes the full response as one finite sequence moves across another, producing a longer output when both inputs have multiple samples. Circular convolution instead keeps a chosen period and folds contributions back into that interval.
This is useful when the underlying signal is genuinely periodic or when an algorithm intentionally uses periodic boundaries. It can also create unexpected image artifacts if a nonperiodic image is filtered through an FFT without suitable padding or boundary treatment.
The difference is mathematical, not an implementation bug. Understanding it prevents wrap-around from being mistaken for real content or for a mysterious failure of the Fourier transform.

How it works

For length N sequences, circular convolution can be written y[n]=sum over m of x[m]h[(n−m) mod N]. The modulo operation wraps indices into the interval from 0 to N−1. In two dimensions, wrapping occurs along both image axes.
Multiplying N-point DFTs and applying the inverse DFT produces the corresponding N-point circular convolution, subject to transform normalization conventions. To obtain full linear convolution of sequences with lengths L and M through FFT multiplication, pad to at least L+M−1 samples. Image filtering also needs a deliberate boundary model, kernel placement and extraction of the desired output region.

Three practical examples

A signal-processing method filters data that are explicitly modeled as periodic.
An FFT-based image pipeline pads inputs to avoid unwanted wrap-around when it needs linear convolution.
A teaching exercise compares spatial convolution with Fourier-domain multiplication using a short sequence.

Think of it this way

Think of a row of samples arranged around a ring. Moving past the last position returns to the first, so contributions can cross the join. Linear convolution instead treats the row as having ends. The analogy captures index wrapping, but actual implementations also need consistent kernel origin and transform scaling to match the intended calculation.

Worked example

Let x=[1,2,3,4] and use the length-four kernel h=[1,1,0,0]. With y[n]=x[n]+x[n−1] and indices taken modulo 4, the output is [5,3,5,7]. The first value is 1+4=5 because the previous sample wraps from the end.
Linear convolution of x with the two-sample kernel [1,1] instead gives [1,3,5,7,4]. Its output has length 4+2−1=5, and the final contribution is not folded into the first position.
Padding both inputs to at least five samples before FFT multiplication can reproduce that full linear result. This example uses a particular indexing convention; shifting the kernel's declared origin shifts the output. The wrap-around distinction remains regardless of how the implementation labels its indices.

Illustrative case study

Fictional case: Cedar Imaging applies a blur by multiplying image and kernel FFTs. A bright object near the right edge creates an unexpected glow at the left edge. The code has used circular convolution on an unpadded image.
The team specifies the intended border behavior, pads appropriately and checks kernel alignment before cropping the result. It compares a small test against a direct spatial implementation. The case illustrates a boundary-condition error, not an inherent defect in FFT filtering. Circular convolution would be correct if periodic wrapping were the intended model for the data.

Common misconceptions and mistakes

Circular convolution is not correlation, although both can be implemented with Fourier methods using different operations. It is not identical to linear convolution unless conditions such as adequate padding are met.
Zero-padding changes how the finite computation is embedded; it does not magically recover information outside an image. Kernel centering and FFT shifts are common sources of alignment errors. The required padding depends on both input and kernel sizes. Finally, spectral leakage from boundary mismatch is related to periodic assumptions but is not simply another name for circular convolution.

Business use

The concept matters in fast image filtering, signal processing and numerical algorithms. Teams may encounter it indirectly when a library uses FFTs for large kernels or repeated convolutions.
Specify whether boundaries should wrap, reflect, replicate or use another extension. Verify the implementation with a small impulse or edge example whose expected output is easy to compute. Record kernel origin, padding, normalization and output cropping. Performance comparisons should use the same mathematical operation: a faster circular result is not an equivalent replacement for a required linear result unless the boundary conditions agree.

Check your understanding

A length-10 sequence is linearly convolved with a length-4 kernel using FFTs. What minimum transform length avoids circular overlap for the full result?

Answer: At least 10+4−1=13, with both inputs padded consistently. A larger convenient transform length may also be used.

Related terms

Discrete Fourier transform
Correlation
Boundary leakage
Linear convolution
FFT
periodic boundary condition
zero-padding
kernel origin

Learn next

Learn linear convolution by hand, then repeat the calculation with modulo indexing. Compare both outputs before using FFT multiplication. In images, test a bright point near a border to make wrap-around and padding choices visible.

Sources

SciPy, Signal Processing tutorial: https://docs.scipy.org/doc/scipy/tutorial/signal.html SciPy, Fourier transforms tutorial source: https://github.com/scipy/scipy/blob/main/doc/source/tutorial/fft.rst Official signal-processing source excerpts consulted. The sequence calculation is original; full convention-specific and editorial review remain pending.

Previous: Circular analysis Next: Circumscription