OE-8.2 Frequency Domain Analysis: DTFT & DFT

What this is and why it exists

Two transforms sit in this unit and they are not interchangeable. One is an analytical tool. The other is what a computer can actually calculate.

The discrete-time Fourier transform is continuous in frequency, so a computer cannot hold it. The discrete Fourier transform is that same thing sampled in frequency, which is exactly what makes it computable.

And there is one trap here that catches nearly everyone. The discrete Fourier transform's native operation is circular convolution, which is not the linear convolution you almost always want.

The vocabulary

• Discrete-time Fourier transform — continuous in frequency, analytical.
• Discrete Fourier transform — sampled in frequency, computable.
• Frequency bin — one of the discrete frequencies the second one reports.
• Periodicity — the assumption that the signal repeats outside the window.
• Circular convolution — convolution that wraps around the ends.
• Linear convolution — ordinary convolution, with no wrap.
• Zero padding — adding zeros so wrapped terms land harmlessly.

The mental model

Start with the relationship, because it makes both transforms make sense at once. The first transform gives a continuous function of frequency — infinitely many values, which no computer stores. Sample that function at evenly spaced points and you have the second. The discrete transform is the continuous one, sampled in frequency.

That sampling has a consequence, and it follows the same logic as sampling in time. Sampling in one domain implies periodicity in the other. Sampling in frequency means the transform assumes your time-domain signal repeats, over and over, outside the window you gave it.

Nobody tells you this and it explains most of the surprises. Suppose the start and end of your window do not match. The assumed repetition then contains a discontinuity that was never in your signal, and the transform faithfully reports its energy across every frequency.

It also explains the convolution trap. Because the signal is assumed periodic, multiplying two transforms and coming back gives convolution that wraps around. Terms that should have fallen off the end reappear at the beginning, corrupting the early samples.

The fix is zero padding, and it is worth understanding rather than memorising. Pad both signals with enough zeros that the full linear result fits inside the window. Then the terms that wrap land in the zero region, where they do no harm, and the circular result equals the linear one. Enough zeros means the two lengths added together, less one.

The properties of both transforms are worth knowing as tools: linearity, shifting, symmetry. They turn many problems into a lookup and an adjustment rather than a computation.

What you should now be able to explain or do

State the relationship between the two transforms in one sentence. Explain what sampling in frequency implies about the time domain. Say why a mismatched window start and end produces energy at frequencies not in the signal. Explain why the computable transform gives circular convolution. Choose a padding length that makes the circular result equal the linear one, and justify it.

Check yourself