The Nyquist limit is a hard ceiling
A signal sampled at 48 kHz has one measurement every 1/48000 second, about 20.83 microseconds. The fastest thing those measurements can describe is one full swing in each sample, which is 24 kHz. That number is Nyquist, and it is not a guideline or a preference. Content above it cannot be represented at all in the sampled stream.
This is why a 96 kHz session is not just a nicer version of 48 kHz. The ceiling moves to 48 kHz and, for a sawtooth with harmonic content proportional to its fundamental, the note range over which harmonics stay below the ceiling rises by roughly one octave. The benefit is real and it is arithmetic.
Why it folds instead of vanishing
The sampling process has an ambiguity. A sample value can be read as belonging to 1 kHz or to 1 kHz plus any exact multiple of the sample rate, because both complete the same number of cycles between samples and land on identical numbers. The sampler has no way to choose, so it delivers the wrong one.
That is the whole mechanism. The offending frequency is mirrored down by the nearest multiple of the sample rate, and it arrives in the range where you are already playing a note, usually as a wrong harmonic of it. The result sounds like grit, growl or a raised pitch, and it moves when you change the note, which is why it is so easy to identify by ear.
| Sample rate | Input frequency | Frequency heard |
|---|---|---|
| 48 kHz | 26 kHz | 22 kHz |
| 48 kHz | 30 kHz | 18 kHz |
| 48 kHz | 60 kHz | 12 kHz |
| 44.1 kHz | 25 kHz | 19.1 kHz |
| 96 kHz | 30 kHz | 30 kHz, no fold |
What 2x, 4x and 8x buy and what they cost
Oversampling runs the generating process at a multiple of the session rate and then filters down to the target rate with a steep low-pass. Because the filter sits well above your actual content, it also removes the aliased copies, which have already been folded far below the wanted signal.
- 2x moves the ceiling to 48 kHz in a 48 kHz project and is the sensible default for most mixing and mastering work.
- 4x roughly doubles the safe note range again and is where most mastering plugins stop.
- 8x is for high notes, heavily saturated waveshapers and any process that generates harmonics above the plain ceiling.
- Cost scales close to linearly with the ratio, so 8x is about four times the CPU of 2x before any filter overhead.
The honest way to describe the gain is that each doubling pushes aliasing products about 6 dB further down relative to your signal. That is a real improvement, but it is an attenuation of a defect rather than the elimination of its cause, and it never becomes zero at any finite ratio.
Band-limited oscillators and polyBLEP
The alternative is to never generate the problem. A naive saw is a ramp that jumps instantly from top to bottom, and that jump is broadband. A fully band-limited saw is computed from the harmonic series with the partials above Nyquist omitted, so nothing ever needs filtering after the fact. It is exact and it is expensive.
PolyBLEP sits between the two. It leaves the naive ramp alone and applies a small polynomial correction to the two samples either side of each discontinuity, shaping the corner into something a true band-limited version approximates. It costs a handful of operations per edge rather than a full transform.
The trade-off is audible in the right measurements. PolyBLEP slightly alters the corner shape, so the harmonic amplitudes of its output do not match an ideal band-limited saw exactly. In practice the difference sits low enough to be a preference rather than a fault, and it is not the excuse people reach for when aliasing is present at a pitch the instrument should never have allowed.
Checking for it, and the myth around it
Play a single sustained note from an oscillator with a spectrum analyser open and watch the partials. Below the aliasing threshold each harmonic sits at a clean multiple of the fundamental. When you cross it, harmonic 5 appears below harmonic 4 and the whole series starts moving backwards. That visual signature is unmistakable.
One claim worth clearing up: aliasing is not a digital-versus-analogue virtue. An analogue oscillator with a hard clipping stage generates harmonics above the tape machine's bandwidth, and those fold too. The difference is that analogue chains fold without a hard mathematical ceiling, which sounds softer and is why the complaint exists at all. That is a description of how the distortion happens, not evidence that one path is more musical.
The bottom line
A frequency above Nyquist does not disappear, it reappears somewhere else, and that is the whole mechanism behind every aliasing complaint. Oversampling moves the ceiling instead of fixing the source, which is why it costs CPU and why a genuinely band-limited oscillator is usually the better fix. Choose 2x as a default, reserve 8x for the top octave, and use a spectrum analyser rather than an opinion to decide.
Confirm it in a blank project
432Hz MASTER leaves a fixed reference pitch, so you can sweep an oscillator upward and hear exactly where the harmonics stop being real.
Frequently asked questions
What exactly is aliasing?
Any frequency component that exceeds half the sample rate appearing in the output at a different, incorrect frequency. A 30 kHz tone sampled at 48 kHz comes back as 18 kHz, because 30 minus 48 is minus 18.
Does oversampling remove aliasing?
It moves the aliasing floor down, typically by 6 dB per doubling, because the aliasing products are pushed further below your signal. It does not remove the problem, and it cannot fix a badly band-limited source without a very large ratio.
Is 4x always better than 2x?
No. Each doubling gives roughly 6 dB more separation between wanted and unwanted content while also doubling the work. 8x is worth it for high notes and for saturated saws; on a bass line at low notes it is usually wasted.
What is polyBLEP?
A correction applied at the discontinuities of a naive waveform to smooth them into something a band-limited version would approximate. It is cheap and sounds good, but it shapes the waveform slightly and does not reproduce every spectral detail of a fully computed band-limited saw.