Spectral Leakage and the Picket-Fence Effect

1. The Root of All Artifacts: Implicit Windowing

When we compute an $N$-point DFT, we are working with a finite sequence of $N$ samples. Conceptually, you can think of this finite sequence as being obtained by taking an infinitely long signal, $x(t)$ or $x[n]$, and multiplying it by a function that is non-zero only for a finite duration. This function is called a window function.

Even if we don't explicitly apply a window, by simply taking $N$ samples and analyzing them, we are implicitly multiplying our signal by a rectangular window. This window has a value of 1 for the duration of our $N$ samples and 0 everywhere else.

This multiplication in the time domain has a profound consequence in the frequency domain. From the convolution theorem, we know that multiplication in one domain is equivalent to convolution in the other. Therefore, the spectrum we compute is not the true spectrum of the original signal, but rather the true spectrum convolved with the Fourier transform of the window function.

This convolution is the single underlying cause of the artifacts we will now analyze. The shape of the window's spectrum dictates the nature of the distortion we see in our DFT output.

2. Spectral Leakage: Energy Spilling Across Frequencies

The Fourier transform of a rectangular window is a sinc function, characterized by a central "main lobe" and a series of decaying "side lobes." When we convolve our signal's true spectrum with this sinc function, the energy from a single frequency component gets "smeared" or "leaked" across the frequency spectrum, following the shape of the sinc.

Spectral leakage is the phenomenon where the energy of a signal at a single frequency spreads out into adjacent frequency bins in the DFT.

This leakage is dictated by the side lobes of the window's spectrum. The higher the side lobes, the more a signal's energy leaks into neighboring frequencies. This is particularly problematic when a strong signal's leakage completely obscures a nearby, weaker signal.

The video below offers an excellent explanation of this process.

The key trade-off, which we will explore in the next lesson, is between main lobe width and side lobe height.

• A narrow main lobe gives high resolution, meaning the ability to distinguish between two closely spaced frequencies.
• Low side lobes result in less spectral leakage, meaning the analysis is more sensitive to weak signals in the presence of strong ones.

The rectangular window has the narrowest possible main lobe (best resolution) but very high side lobes (worst leakage).

3. The Picket-Fence Effect: What You Can't See

The second artifact, the picket-fence effect, arises from the DFT's discrete sampling in the frequency domain. The DFT calculates the spectrum at $N$ discrete frequency "bins." The true spectrum, however, is continuous (it's the DTFT).

Imagine trying to view a continuous mountain range by peering through the gaps in a picket fence. You can only see what's directly in front of the gaps. If a peak of the mountain range falls behind one of the fence posts, you might miss it entirely or underestimate its height.

This is exactly what happens with the DFT:

• If a signal's true frequency peak falls exactly on a DFT bin, you will measure its correct amplitude.
• If the peak falls between two DFT bins, its energy will be distributed between those two bins, and the measured amplitude in both bins will be lower than the true peak amplitude. This reduction in peak amplitude is often called scalloping loss.

For a more formal and practical definition, especially relevant to your background in electrical engineering, let's turn to an application note from Texas Instruments.

Mitigating the Picket-Fence Effect with Zero-Padding

As the TI note mentions, we can reduce the picket-fence effect by adding zeros to the end of our time-domain signal before computing the DFT. This is called zero-padding.

Zero-padding to a total length of $M>N$ computes an $M$-point DFT. This doesn't add any new information to the signal, but it results in the DFT being calculated on a finer grid of $M$ frequency points instead of $N$. In our analogy, this is like adding more, narrower gaps to our picket fence, giving us a better view of the landscape behind it.

Crucially, zero-padding does not improve resolution. Resolution (the ability to separate two close frequencies) is determined by the length of the original, non-padded time-domain window ($N$). Zero-padding simply interpolates the spectrum, giving a smoother-looking plot and a more accurate estimate of the peak amplitudes and frequencies that were already resolved.

The following video provides an excellent visual demonstration of this concept.

4. The Special Case: When Leakage Disappears

You might have noticed that sometimes, when analyzing a pure sinusoid, the DFT shows a single, sharp peak with no apparent leakage. This happens under a very specific condition: when the signal completes an integer number of cycles within the N-sample window.

Mathematically, this occurs when the signal frequency $f_{sig}$ is an exact multiple of the DFT's frequency resolution, $f_s/N$.
$f_{sig}=k\cdot \frac{f_s}{N}\text{for some integer }k$

In this case, the implicit periodic extension that the DFT assumes is continuous. In the frequency domain, a remarkable alignment occurs: the peak of the sinc function (from the rectangular window's spectrum) lands exactly on one DFT bin, and all other DFT bins land exactly on the zero-crossings of the sinc function. Thus, we don't "see" the side lobes, and there is no leakage.

The moment the signal's frequency deviates even slightly from being a multiple of $f_s/N$, the DFT bins no longer align with the sinc zero-crossings, and the side lobes become visible as leakage.

This is visually explained in Figure 1 of the article from LINK.

This figure from the article "Understanding the DFT..." illustrates the effect of windowing on a single sinusoid. In case A, an integer number of periods is sampled. The DFT frequency bins (blue dots) align perfectly, with only the central peak being non-zero. In case B, a non-integer number of periods is sampled. The bins no longer fall on the sinc function's zero-crossings, and we observe spectral leakage.

The TI application note (LINK) refers to this ideal condition as synchronous sampling. In fields like power systems analysis, significant effort is made to achieve synchronous sampling to get the most accurate phasor measurements, free from leakage artifacts.

Conclusion

In this lesson, we have demystified two of the most common artifacts in practical spectral analysis. You now understand that these are not arbitrary errors but predictable consequences of how the DFT works.

Key Takeaways:

• The DFT implicitly uses a rectangular window, which means the computed spectrum is the true spectrum convolved with a sinc function.
• Spectral Leakage is the smearing of energy into adjacent frequency bins, caused by the side lobes of the window's spectrum. It obscures weak signals.
• The Picket-Fence Effect is the potential to mis-measure a signal's amplitude and frequency because its peak falls between the DFT's discrete frequency bins.
• Leakage is a consequence of the finite window length (truncation). The picket-fence effect is a consequence of the discrete sampling in the frequency domain.
• These effects are minimized when the signal is periodic within the window (synchronous sampling). Leakage is managed with different window functions, and the picket-fence effect is mitigated by zero-padding.

Preview of the Next Lesson:

We have seen that the rectangular window provides the best resolution but suffers from severe spectral leakage. This suggests a trade-off. Can we find other window functions that offer a better compromise? In our next lesson, we will do just that, as we move to apply various window functions (e.g., Hamming, Hanning) to reduce spectral leakage and analyze their trade-offs.