Is every convolution with a fixed kernel LSI?

On the usual unrestricted domain, yes; finite-image boundary and output conventions need separate attention.

Does shift invariance mean the output does not move?

No. The output moves with the input while preserving the same response pattern.

Can a camera's blur be LSI only approximately?

Yes. A single kernel may fit a limited region or operating condition while failing elsewhere.

AI Glossary / Computer vision and multimodal AI / Intermediate

Linear shift-invariant (LSI) filter

Definition

A linear shift-invariant filter, or LSI filter, preserves superposition and responds the same way when an input pattern is shifted in space. For ordinary discrete images on an appropriate domain, it can be described by convolution with a fixed impulse response or kernel. The same pattern produces the same filtered pattern at a different location, subject to boundary effects.

What it means

The name combines two independent properties. Linearity means that scaled and added inputs produce correspondingly scaled and added outputs. Shift invariance means that translating an input translates its output without otherwise changing the response.

In image processing, this often means applying the same fixed neighborhood weights everywhere. A point of light at one location is spread according to a particular pattern; the same point elsewhere produces a translated copy of that pattern.

This is a model, not a guarantee about every camera or filter. Blur may change across a lens's field of view, and an adaptive filter may respond differently depending on surrounding intensities. Such behavior can violate shift invariance, linearity, or both.

How it works

Determine the response to a unit impulse: one sample with value one and zero elsewhere. This response is the kernel that characterizes an LSI system under the usual convolution assumptions.

Any discrete input can be viewed as a sum of shifted, scaled impulses. Linearity allows their responses to be added, and shift invariance says that each response is the same kernel shifted to the impulse's location. Summing them produces convolution.

The Fourier transform provides another description: convolution in the spatial domain corresponds to multiplication by a frequency response under compatible transform and boundary conventions. Different spatial frequencies can therefore be attenuated or amplified predictably.

Finite images need an extension or border rule. Cropping, padding, and image boundaries can make a practical operation fail an unrestricted translation test even when its interior uses a fixed convolution kernel.

Three practical examples

1. Controlled blur modeling: An imaging engineer approximates uniform defocus over a limited field using one point-spread function and evaluates how well the approximation holds.

2. Feature preprocessing: A vision pipeline applies a fixed Gaussian or derivative kernel across an image so its interior response is consistent across location.

3. Frequency analysis: A researcher studies how a filter suppresses rapid intensity oscillations by examining its frequency response before testing it on photographs.

Think of it this way

Think of using the same rubber stamp wherever a dot appears on a sheet, scaling the ink contribution according to each dot's strength and adding overlapping impressions. The analogy captures the impulse-response construction. It is limited because real ink can saturate or interact, whereas the mathematical LSI model requires exact addition without those nonlinear effects.

Worked example

Consider a one-dimensional kernel with values 0.25, 0.5, and 0.25 at offsets −1, 0, and +1. An impulse of amplitude 8 at position 5 produces output values 2, 4, and 2 at positions 4, 5, and 6.

Move that input impulse to position 9. Shift invariance predicts the same output values at positions 8, 9, and 10. The response moves four positions but keeps its shape and amplitude.

Now add an impulse of amplitude 4 at position 6 to the original impulse at position 5. Its contribution is 1, 2, and 1 at positions 5, 6, and 7. Adding the two responses gives output values 2, 5, 4, and 1 at positions 4 through 7.

Their sum is 12, matching the combined input amplitude because this kernel sums to one. That preservation follows from the chosen kernel and full output support; it is not the definition of LSI. Clipping the result or discarding part of it at an image boundary would change the calculation.

Illustrative case study

Fictional case study: A scanning team models its instrument's blur using a single kernel measured near the image center. Reconstruction looks reasonable centrally but becomes less reliable near the corners.

The team measures point-like targets at several positions and finds that the response changes across the field. The instrument is not well described by one shift-invariant blur over the whole image.

It limits the simple model to a region where validation supports it and investigates a spatially varying alternative for wider coverage. The team reports the applicable region and residual errors. The scenario illustrates checking an assumption rather than concluding that an elegant convolution model must describe the entire device.

Common misconceptions and mistakes

Linear does not automatically mean shift-invariant. Fixed position-dependent gains can be linear while responding differently at different locations.

Shift-invariant does not automatically mean linear. Some neighborhood operations treat translated patterns consistently but fail superposition.

A convolution kernel does not have to sum to one. Smoothing, differentiation, and sharpening can use different coefficient sums and frequency responses.

Applying a Fourier transform does not remove boundary assumptions. Circular convolution and zero-padded linear convolution are different operations unless handled appropriately.

The term invariant does not mean that an output ignores object position. Shifting the input shifts the output; the response rule stays the same. This is often called translation equivariance in machine-learning discussions.

Business use

LSI models are useful because one impulse response or frequency response can summarize a system's behavior. This supports filter design, blur characterization, efficient implementation, and controlled comparisons between processing pipelines.

Validate the model across the relevant image area and intensity range. Test whether point responses change with location and whether scaled inputs remain proportionally scaled. Document padding and cropping rules so that edge behavior is reproducible. If the application contains strong spatial variation or adaptive behavior, use an appropriate richer model rather than forcing all observations into one kernel.

Check your understanding

An impulse shifted three pixels right produces a differently shaped response in the image interior. Is that consistent with an exact LSI model?

Answer: No. An exact shift-invariant response would move three pixels right without changing shape.

Related terms

Linear filter
Convolution
Point spread function
impulse response
frequency response
translation equivariance
spatially varying blur

Learn next

Construct convolution by adding shifted impulse responses, then compare that result with direct filtering. Test the same operation near finite-image borders and identify which assumptions change there.

Convolution

Sources

Steven W. Smith, Linear Systems: https://www.analog.com/media/en/technical-documentation/dsp-book/dsp_book_Ch5.pdf

Steven W. Smith, Convolution: https://www.dspguide.com/CH6.PDF

Both primary book chapters were opened. Impulse calculations and the scanning scenario are original teaching examples.

Previous: Linear separator Next: Linear temporal logic