Motivation
A camera converts a continuously varying light field into a finite grid of measurements. Fine texture can become a false coarser pattern when the sampling grid cannot distinguish its variation. The camera-warping chapter cautioned that downsampling needs a suitable prefilter; this unit derives a concrete aliasing example and makes a finite recovery condition explicit.
Sampling can identify different signals
Definition 9.1.1
(Point sampling and frequency)
Let be a sampling rate in samples per unit distance. Under the ideal point model, a continuous one-dimensional signal produces samples at integer . For a sinusoid with spatial frequency cycles per unit distance, its complex samples are . Two different signals that produce the same samples are called aliases of each other.
Proposition 9.1.2
(Aliases of a sampled sinusoid)
For every integer , the frequencies and have the same complex samples. For real cosines, and also have the same samples.
Proof
(Periodicity of the exponential and of the cosine)
The samples of frequency are , because for integers . For the cosine,
since cosine is even and -periodic.
These equalities concern the stated point-sampling grid. A finite field of view, noise, pixel averaging and nonuniform sample positions introduce additional effects; no formula here says they reveal all missing information.
Counterexample mode
A high frequency becomes indistinguishable from a lower one
Sample at samples per millimeter. The distinct signals and give identical values at every , because for integer . At they are different. Samples alone cannot tell which of these two signals was present. Adding the same constant intensity and scaling both signals preserves the ambiguity.
Worked example 9.1.3
(The apparent frequency of several inputs)
Let samples per unit and sample . Reduce modulo , and if the result exceeds replace it by minus the result. This gives apparent frequency for , and , and apparent frequency for and . Indeed . The frequency is its own fold point: its cosine samples are , the fastest alternation the grid can show. All five inputs and therefore collapse onto only two distinct sample patterns.
A finite periodic recovery result
Theorem 9.1.4
(Recovering a restricted Fourier family)
Let be an integer, and suppose the unknown periodic signal is known in advance to have the finite form
If equally spaced samples are known, each coefficient in this declared family is uniquely recovered by
Proof
(Orthogonality of roots of unity)
Substitute the finite sum for and interchange the finite sums:
The inner geometric sum is when is divisible by and otherwise. Since , divisibility occurs only for , leaving . Because the formula returns from the samples, two signals of the family with the same samples have the same coefficients.
Proof X-ray
Why the number of samples matters
The hypothesis is used at exactly one point: to guarantee , so that among the indices no two different ones differ by a multiple of . The key move is the orthogonality of the -th roots of unity, which turns a sum over samples into a coefficient extraction.
When two exponentials of the family can agree on the grid, since depends only on modulo . This is the aliasing of the previous section in coefficient form. The argument proves uniqueness within the stated finite family; it does not identify an arbitrary continuous image from finitely many samples. Equality at a boundary frequency needs care: distinct positive and negative frequencies can coincide on a grid.
Worked example 9.1.5
(Too few samples can hide a variation)
For and , the sufficient condition fails. At , the nonzero function is sampled as zero at every index, just like the zero function. Here the positive and negative frequency-two components are indistinguishable on the grid. The example shows why a claim of unique recovery needs its class and sampling assumptions.
Common mistake
(Sampling at exactly twice the frequency does not always suffice)
With and , the signal has samples for every , so it cannot be told apart from the zero signal, whereas gives . At a pure cosine is still visible, but a sine component of the same frequency is lost entirely. The recovery condition has a strict inequality in the same spirit.
Averaging before sampling
A real sensor pixel measures over a nonzero footprint. In one dimension, an ideal uniform footprint of width can be modeled by
Proposition 9.1.6
(Response of a box footprint to a cosine)
For with ,
Proof
(Integrating over a centered interval)
Let . Then
By the identity with and , the numerator is . Dividing by gives .
At the continuous limiting factor is . The footprint is a prefilter: its frequency response changes the amplitudes before sampling. It may attenuate fine structure, but a box average does not eliminate every frequency that can alias.
Worked example 9.1.7
(A footprint attenuates but does not remove an alias)
Take samples/mm and mm. For the two earlier frequencies, the factors are and . The higher frequency is more attenuated in this example but its factor is nonzero. Its samples still follow the same discrete cosine pattern as frequency , scaled by a different factor; without an amplitude assumption this does not uniquely identify the underlying frequency.
Practical interpretation
In two dimensions, a rectangular periodic grid aliases spatial frequencies according to the corresponding horizontal and vertical sampling periods. A prefilter should be selected for the footprint and target sampling grid before downsampling. Real lens blur, sensor integration, finite boundaries and spatially varying warps need their own models. Interpolation constructs values between existing samples; it does not restore arbitrary frequencies already lost to sampling. A sharper-looking output is not by itself evidence that the missing scene detail was recovered.
Common mistake
(Filtering after sampling cannot undo aliasing)
In the counterexample above, and have identical samples at . Any procedure that uses only those samples, such as a sharpening or smoothing filter applied after downsampling, returns the same output for as for , so it cannot restore the one that was sampled. The prefilter must act before sampling, on the continuous signal or at a finer grid.
Exercises
Quick check
(If samples per unit, which lower nonnegative frequency below has the same cosine samples as frequency ?)
Use the reflection .
Solution · Answer
Frequency , because and the cosine samples agree at all integer indices.
Quick check
(For the periodic family above, does satisfy the stated condition for ? What happens to the two exponentials with frequencies and when ?)
Compare with , and evaluate the exponentials on the grid.
Solution · Answer
Yes: . At , both exponentials take the values on the sample grid, so their coefficients cannot be separated by those samples alone.
Quick check
(For footprint width and frequency , compute .)
Substitute .
Solution · Answer
.
Quick check
(Why does a finite set of samples not uniquely determine an arbitrary continuous signal even if a particular restricted-family theorem permits recovery?)
Which assumption does the theorem make about the unknown signal?
Solution · Answer
The theorem assumes a finite Fourier family known in advance. Other continuous signals outside that family can agree at all the sampled positions, as the preceding counterexample demonstrates.
Further study
General sampling theorems, two-dimensional frequency geometry, sensor point-spread functions, reconstruction filters and adaptive antialiasing require additional analysis. This unit does not implement an image resampler or certify its visual quality.