A continuous signal becomes a sequence
Sampling a continuous signal at rate records its values at times , where is an integer. For , the resulting sequence is
Only these values are retained. The sequence alone cannot reveal what happened between samples without additional assumptions.
Definition 4.1.1
(Aliases at a sampling rate)
Fix . Two cosines and are aliases at rate if their sampled sequences agree at every integer :
Aliasing is a statement about the sequences only; the two continuous signals may be different functions.
Why two frequencies can agree
For any integer , replacing by changes the sampled angle by , an integer number of full turns. Therefore the samples are identical for every integer , with the same phase. Cosine is even, so a sign reversal also produces an alias if the phase is reversed:
The phase qualification matters. Negating frequency while keeping a general phase unchanged need not give the same sequence. These identities show ambiguity, not a claim that every pair of signals is an alias. In precise form:
Theorem 4.1.2
(Aliasing identities for sampled cosines)
Let , and let be real numbers. For every integer :
- is an alias of at rate , with the same phase.
- is an alias of at rate : the frequency and the phase change sign together.
Proof
(Whole turns and the evenness of cosine)
For part 1, write . The sampled angle for is . Because and are integers, is an integer, so is a whole number of turns and for every integer .
For part 2, the sampled angle for with phase is , and because cosine is even.
Proof X-ray
Which hypotheses carry the proof
The proof compares angles, not values: it shows that the two angles differ by a multiple of at every sample. Two integrality hypotheses are used and neither can be dropped. The index is an integer because the signal is read only at . The shift must be an integer for to be a multiple of for every . For , and the shifted frequency is Hz; at the angles give and , which differ because . So Hz is not an alias of Hz at this rate.
The argument closes without any limit or approximation, because it needs only finitely many facts per sample. It also proves less than it may seem to: at a time that is not of the form the angles differ by , which need not be a multiple of . For instance, at s with , and , the two signals have values and . Equality holds only on the sampling grid.
The second identity needs a further hypothesis: the phase must be reversed together with the frequency.
Counterexample mode
Reversing the frequency without reversing the phase
It is tempting to claim that always has the same samples as . Take , and , and look at . The signal with gives . With and the unchanged phase one gets , so the claim fails. Reversing the phase as well gives , which agrees. The claim does hold when or , because then and differ by a multiple of (this is the situation in the zero-phase example below), but it fails for a general phase.
Worked example
Worked example 4.1.3
(A 2500 Hz cosine sampled at 3200 Hz)
Take samples per second and . Since ,
Thus a zero-phase 700 Hz cosine gives exactly the same samples as the 2500 Hz cosine. Their values between samples need not agree. The first three samples are , and . Neither sampled sequence, by itself, identifies which continuous frequency was present.
With a phase , the 2500 Hz samples instead equal those of a 700 Hz cosine with phase . For instance, at both phases give the same cosine, but this equality at one sample does not justify keeping the phase unchanged at every sample.
Worked example 4.1.4
(A phase carried through an alias)
Let and . Since , the identities give the alias , with the phase reversed. Check two samples. At the original angle is , so the sample is ; the alias gives and the same value. At the original angle is , whose cosine is , and the alias gives . Had the phase been kept at , the Hz cosine would give at , which is wrong.
Worked example 4.1.5
(Two extreme cases: a multiple of the rate and half the rate)
Let . First, has samples for every : it is an alias of the constant signal with frequency , taking and . Second, has samples for every , so it is indistinguishable from silence, while has samples . At the frequency the recorded amplitude depends on the phase, and a sine of that frequency vanishes entirely at the samples. This is why the sampling condition below is a strict inequality.
Avoiding a false conclusion
In the standard bandlimited model, if a signal is known to have no frequency content with absolute frequency at or above a positive cutoff , the strict condition is sufficient for ideal reconstruction under the sampling theorem’s usual assumptions; other prior information can allow other approaches. This condition places that band below half the sampling rate. If the input contains frequency content that would alias, suitable filtering before sampling is needed to suppress that content: once aliases overlap in the recorded samples, a later digital calculation cannot infer which original signal produced them without further information. The strict inequality avoids treating the special endpoint as an ordinary pair of distinguishable frequencies.
Exercises and solutions
Quick check
(At , find an alias between 0 and 500 Hz for a zero-phase 1300 Hz cosine.)
Use part 1 of the identities with an integer shift.
Solution · Answer
Subtract : , and lies between and . The alias is the zero-phase Hz cosine.
Quick check
(At , identify the phase of the 200 Hz alias of .)
Write as .
Solution · Answer
Shifting by one rate, the Hz cosine has the same samples as the Hz cosine with the same phase . By part 2, that is the Hz cosine with phase reversed. The phase is .
Quick check
(Does filtering after sampling distinguish two signals that already gave identical samples?)
Think about what the digital operation receives as input.
Solution · Answer
No. The same deterministic digital operation receives the same sequence in both cases, so it produces the same output in both cases.
Quick check
(If a known cutoff is Hz, does a sampling rate of 800 Hz satisfy the stated strict condition?)
Compare with .
Solution · Answer
No. The condition is , and is not greater than ; the rate must be more than 800 samples per second.