A sampled function of time on the left, with its spectrum next to it. Both displayed as periodic.
Logically, neither signal nor spectrum precedes the other, or are even distinguishable. For introductory purposes the remarks that follow take an applied perspective. A physical measurement takes place in time and is bounded in duration. The resulting spectrum can be used to reconstruct the signal beyond the measured samples, and yields a periodic function of time — with possible “glitches” at the boundaries, which have become implicitly linked. The next section presents some fundamental examples.
A general discussion of samples and sequences would include these talking points:
By contrast, this section takes a stereotypical (verging on dogmatic) laboratory perspective:
We choose the number of data points N as 256 for adequate screen resolution. N is dimensionless: the DFT does not come equipped with units. Its functions are defined on the bare integers only.
A real-world measurement will most surely have units. A Hertz is no “smaller” or “larger” than a second, but the coincidence of the sha function Ш in both domains is important. They are the first two examples above. The criterion that
implies that
which in turn yields the calibration described below.
It is easy to assume that ΔF = 1 / ΔT 🚫 and ΣF = 1 / ΣT 🚫 but neither is true.
ΣT = 𝚺1..NΔT = N∙ΔTso ΔT is now determined as:
It is physically intuitive to treat time as one-sided, running from zero (start of measurement) to ΣT (end of measurement) in steps of ΔT.
Human-readable displays usually show time as one-sided. There is no requirement that the last half of the samples be zero. The display above starts the signal at time zero, and draws more than two full measurement durations.
However...you will find when entering custom functions that it does indeed matter which range you use. Try this experiment.
At this point the we switch over to the other column for steps 3 and 4. The half-values are so directly dependent on the full values that they come as afterthoughts.
This is primarily just for symmetry in the equations.
One is tempted to call this the “Nyquist” time 𝑇𝑁, but it depends on how you want the analogy to work. If you want it to mean the time beyond which the signal must be zero,” then 𝑇𝑁 should be the measurement duration ΣT = 16 s.
When drawing a sampled (hence implicitly periodic) time domain as literally circular, it is natural to treat the “last” samples as being in negative time, although events in the “negative half” did not really precede events in the positive half.
The reciprocal of 𝑇s from step 2 must be 𝐹s:
ΣF = 𝚺1..NΔF = N∙ΔFso ΔF is derived as:
It is less plausible to treat frequency as one-sided, running from zero (DC) to ΣF.
Human-readable spectra usually omit the redundant bins, or display them as two-sided. When the signals are real, the spectra have even magnitude. The display above centers the frequency at zero, and draws more than two full measurement durations.
This is important enough for its own name and symbol
It is a prerequisite for unambiguous sampling that a signal have no components above 𝐹𝑁. The spectrum values that a DFT provides in the “high half” do not represent beyond-Nyquist frequencies, but negative frequencies paired with positive ones. The information in N real samples is the same as in N/2 complex spectral values.
This apparent asymmetry between the domains’ internal symmetry can be avoided by using the Cosine or Hartley transform.
Drawing a sampled (implicitly periodic) frequency domain as literally circular poses fewer conceptual blocks, because the idea of negative frequency does not seem to raise issues of causality.