Does translation preserve Fourier magnitude?

Under the theorem's ideal or circular-shift conditions, yes.

Does unchanged magnitude uniquely reveal the shift?

No. Phase and additional conditions are needed, and ambiguities can remain.

Is array frequency-centering the same operation?

No. Reordering displayed frequency bins differs from translating the underlying signal.

AI Glossary / Computer vision and multimodal AI / Advanced

Shift theorem

Definition

The Fourier shift theorem describes how translating a signal changes its Fourier transform. Under the convention F(ν)=∫f(x)e^(−i2πνx)dx, translating f to g(x)=f(x−a) gives G(ν)=e^(−i2πνa)F(ν). The translation changes phase while preserving magnitude. For a finite discrete Fourier transform, the corresponding exact integer-shift statement concerns circular shifts of the sampled sequence.

What it means

Moving an ideal signal does not change which frequency magnitudes it contains, but it changes how those components align in space or time. That alignment is encoded in phase. The theorem helps explain why magnitude alone loses translation information. However, moving an object within a fixed photograph can also change cropping, visibility or illumination. Such a new image is not necessarily an exact translation of the old one, so the ideal relationship should not be applied without considering what the capture actually changed.

How it works

Substitute the translated coordinate into the Fourier integral. The translation contributes a complex exponential factor whose magnitude is one and whose phase varies linearly with frequency. In two dimensions, the factor depends on both spatial-frequency components and both displacement components. For an N-sample DFT, a circular shift by a samples multiplies frequency bin k by e^(−i2πka/N) under the stated negative-exponent convention. Sign conventions must match the transform definition. Finite-window cropping, padding and interpolation can introduce effects beyond this exact relationship.

Three practical examples

An image-registration exercise uses phase differences to estimate a translation between suitably related images.
A signal-processing lesson compares the spectra of an impulse before and after a circular shift.
A debugging check verifies that a frequency-domain translation uses the intended sign and boundary convention.

Think of it this way

Imagine a group of musicians playing the same repeated pattern but starting together at a later time. The component rhythms remain the same, while their timing relative to the original clock changes. Fourier phase records an analogous alignment change. The analogy is limited: the exact theorem assumes a mathematical translation, so cutting off the beginning or ending of a recording introduces an additional windowing change rather than merely delaying an unlimited signal.

Worked example

In an original four-sample DFT illustration, x=[1,0,0,0] has transform X=[1,1,1,1] under the negative-exponent convention. Circularly shifting the impulse one sample to the right gives y=[0,1,0,0]. The shift factor across bins k=0,1,2,3 is [1,−i,−1,i], so Y=[1,−i,−1,i]. Every magnitude remains 1, but the phases differ. This example makes the circular boundary convention explicit. A shift implemented by inserting zeros and discarding samples from a general finite sequence need not preserve the same spectrum because it can remove original signal content.

Illustrative case study

A fictional imaging team implements a frequency-domain alignment step. A synthetic circular-shift test works exactly, but real camera images show residual errors near borders. The team investigates cropping, newly visible scene content and interpolation rather than concluding that the theorem has failed. It evaluates alignment on overlapping regions and documents the finite-image assumptions. This scenario illustrates the difference between an exact mathematical property and the conditions of a practical registration task. It does not claim that phase information alone resolves rotation, depth-dependent motion or every ambiguity in repeated image patterns.

Common misconceptions and mistakes

The shift theorem does not say that phase is unchanged. It says magnitude is preserved under the stated ideal translation while phase acquires a frequency-dependent factor. Another mistake is confusing spatial translation with moving the displayed frequency origin using an array-reordering operation. Circular shift and cropped translation are also different. A phase ramp's sign depends on both transform and shift conventions, so memorizing a sign without defining them invites errors. Finally, equal Fourier magnitudes do not establish that two arbitrary signals are merely translated copies of one another.

Business use

The theorem supports alignment, signal analysis and efficient image-processing operations. Use it with clearly defined coordinates, transform conventions and boundary handling. Verify implementations on synthetic translations with known answers before evaluating realistic captures. For operational image alignment, assess overlap, interpolation error and deviations from pure translation. Preserve the distinction between a correct spectral calculation and successful registration of real content. The practical benefit comes from exploiting a precise mathematical relationship while testing whether the data satisfy the assumptions that make the relationship useful.

Check your understanding

What was the second bin of the shifted four-sample transform?

Answer: It was −i under the stated convention.

Did its magnitude change from 1?

Answer: No.

Why were real cropped images qualified separately?

Answer: A cropped shift can add or discard content, so it is not necessarily the same operation as an exact circular translation.

Related terms

Phase correlation
Image registration
Discrete Fourier transform
Fourier phase
circular shift
windowing

Learn next

Compute the four-sample example directly from the DFT definition. Reverse the shift and check the phase-factor sign. Then shift a nontrivial finite sequence using both circular wrapping and zero-filled cropping, comparing magnitudes to identify exactly which assumption changed.

Sources

MIT, Foundations of Computer Vision, Fourier Analysis: https://visionbook.mit.edu/image_processing_fourier.html The four-sample calculation, musician-timing analogy and imaging-team case are original educational illustrations. The transform convention and circular-shift scope are explicit, avoiding a claim that arbitrary cropped photographs must share identical Fourier magnitudes.

Previous: Shift-reduce parsing Next: Shingling