2. Sampling Theorem - Digital Audio Fundamentals

Main ideas / concepts

• Why start with a sine wave

  • A continuous-time sine wave has infinitely many possible values over time.
  • For digital audio, we can’t store infinitely many values, so we convert the signal into discrete samples taken at specific times.

• Analog → Digital conversion as sampling

  • The key question becomes: how often should we measure the signal?
  • Sampling is described as taking a fixed number of samples per second (the sample rate).
  • Example given: sampling a 1 Hz sine wave
    • 30 samples per second (sample every 0.03 s)
    • other example rates like 20, 8, and 44100 samples per second

• Nyquist–Shannon Sampling Theorem (core rule)

  • Any band-limited continuous-time signal can be converted to/from digital signals without loss if sampled at a rate:
    • at least twice the highest frequency component
  • Two key conditions:
    1. Sampling rate must be high enough
       • For a signal containing a 10 kHz component, sampling rate must be > 20 kHz.
       • For a 1 Hz signal, the theoretical minimum discussed is 2 samples per cycle, but the video illustrates with at least ~3 points as a “minimum required” example.
    2. The signal must be band-limited
       • The input must contain no frequency components above the allowed maximum.
       • This requires filtering (low-pass) so frequencies above the limit are removed before sampling/reconstruction.

• Misconception addressed: “Higher sample rate always sounds better”

  • Mathematically, if the signal truly is band-limited and the theorem conditions are met, then:
    • a low sampling rate that satisfies Nyquist and a higher sampling rate can reconstruct to the same indistinguishable analog signal.
  • So there’s no inherent quality gain merely from having more samples if the bandwidth is already sufficient.

• Practical reality: anti-aliasing filters aren’t perfect

  • Real low-pass filters can’t cut off abruptly at exactly the Nyquist frequency.
  • Their transition is gradual, so unwanted high-frequency components may “spill over,” causing aliasing.

• Nyquist frequency and aliasing

  • The maximum frequency that can be represented without aliasing is the Nyquist frequency (half the sampling rate).
  • If frequencies above the Nyquist frequency are present, they fold back into the allowed spectrum, creating artifacts.
  • The video links this to amplitude modulation/interference-like effects.

• Engineering Nyquist theorem (practical adjustment)

  • Because of imperfect filters and hardware limitations, the video suggests a safer rule of thumb:
    • instead of sampling at 2× the highest representable frequency,
    • sample at about 2.5× the highest representable frequency.

• Demonstrations using Audacity

  • The video uses Audacity to visualize and listen to sampled sine waves at different rates to show:
    • correct reproduction when within band limits
    • failure/odd behavior at or near Nyquist
    • improvement when the sampling rate increases (with sufficient “buffer”)

---

Methodology / instructional points (as presented)

1. Choose a sampling rate for your target bandwidth

   • Determine the highest frequency component you want to represent: ( f_{max} ).
   • Theoretical Nyquist condition: sample at ≥ 2× ( f_{max} ).
   • Practical recommendation (engineering Nyquist): sample at about ≥ 2.5× ( f_{max} ) to tolerate real filter roll-off.

2. Ensure the signal is band-limited before sampling

   • Apply a low-pass filter so the signal contains no energy above the Nyquist frequency.
   • Lossless reconstruction requires the absence of higher frequencies.

3. Use the Nyquist frequency as the cutoff reference

   • Nyquist frequency = ( \text{sampling rate} / 2 ).
   • In the theoretical model, anti-aliasing/anti-imaging filtering ensures reconstruction contains only allowed frequencies.

4. Understand the consequence of filter imperfections

   • Real filters roll off gradually, so frequencies near/above Nyquist may remain.
   • These components cause aliasing, which can manifest as:
     • amplitude modulation
     • incorrect spectral content
     • interference-like behavior (later described via a frequency-pair “beating” test)

5. Validate practically (example workflow concept)

   • In Audacity (as used by the video):
     • set a low sample rate to test limits
     • generate sine tones at/near Nyquist
     • plot spectrum / observe waveform display issues
     • increase sample rate to confirm correct reproduction when sufficiently below Nyquist with buffer

---

Key examples / outcomes described

• 8 kHz sample rate (Nyquist = 4 kHz)

  • 1 kHz tone: works well (sounds right; spectrum shows 1 kHz).
  • 4 kHz tone (exactly Nyquist):
    • appears “nulled out” (no visible signal),
    • but can be brought out by changing phase—illustrating an edge-case behavior.
  • 3,999 Hz tone:
    • within Nyquist, but still shows bizarre amplitude modulation
    • explained as aliasing caused by practical filter/spillover limitations.

• Increasing sample rate to 11,025 Hz

  • 3,999 Hz tone: reproduced more correctly (no weird amplitude modulation).
  • Reason: greater buffer between ( 3,999 ) Hz and the Nyquist cutoff.

• Further sample rate increases (44,100 Hz, 96,000 Hz, 192,000 Hz)

  • Audibly: no perceptible difference for that particular 3,999 Hz test.
  • Visually: waveform/resolution differences in display, but reconstructed audio quality (for this case) is essentially the same.

• Aliasing demonstrated via frequency “beating”

  • Create tones at 400 Hz and 401 Hz, combine them:
    • audible “beating” occurs due to interference between slightly different frequencies.
  • Used as an analogy for how aliasing can cause interfering components that change amplitude over time.

---

Speakers / sources featured

• Narrator / video speaker (unnamed)
• Monty from Xiph.org (referenced as the source of an external demonstration video)
• Audacity (software used for demonstrations; brand referenced, not a person)