When a saturation or distortion plugin sounds unexpectedly brittle or "digital," the problem is often not the algorithm itself. It is aliasing: a specific, mathematical side effect of nonlinear processing in the digital domain. Understanding it precisely changes how you think about every waveshaper, tube emulator, and distortion pedal plugin in your session.
Why Nonlinear Processors Create New Frequencies
A linear processor, like a shelving EQ or a compressor, modifies the amplitude or phase of frequencies already present in the signal. It does not create new ones. A nonlinear processor does. When you apply soft clipping to a sine wave, the output is no longer a pure sine, it is a combination of the fundamental frequency plus its harmonics at integer multiples: 2x, 3x, 4x, and so on.
In analog processing this is unproblematic. Harmonics simply exist at whatever frequency they fall, and if they are above the range of human hearing, they contribute nothing audible. In the digital domain, a hard ceiling exists at the Nyquist frequency, half the sample rate. At 44.1 kHz, that ceiling is 22.05 kHz. Any signal component above this limit cannot be represented. Instead of disappearing, it folds back into the audible spectrum as an alias at a different, unrelated frequency.
The Aliasing Calculation
The fold-back formula is straightforward. For a harmonic at frequency f processed at sample rate fs, if f exceeds the Nyquist frequency (fs/2), the aliased frequency appearing in the output is:
f_alias = fs - f (for fs/2 < f < fs)For frequencies above fs, the same folding repeats periodically.
A concrete example: an 8 kHz input signal through a hard clipper at 44.1 kHz. The clipping generates harmonics at 8, 16, 24, 32, 40, 48 kHz and beyond.
- 8 kHz: below Nyquist, audible, clean
- 16 kHz: below Nyquist, audible, clean
- 24 kHz: above Nyquist (22.05 kHz), alias at 44.1 - 24 = 20.1 kHz
- 32 kHz: alias at 44.1 - 32 = 12.1 kHz
- 40 kHz: alias at 44.1 - 40 = 4.1 kHz
- 48 kHz: 48 - 44.1 = 3.9 kHz (wraps once), alias at 3.9 kHz
The alias frequencies 20.1, 12.1, 4.1, and 3.9 kHz are not part of the harmonic series of 8 kHz. They are inharmonic. Their pitch relationship to the original note changes with every note you play, producing the characteristic digital buzzing that no amount of mixing can fully fix.
How Oversampling Eliminates the Problem
Oversampling solves this by processing at a higher sample rate, then filtering and decimating back down to the original rate. The four steps are:
- Upsample: Insert interpolated samples to raise the effective sample rate. At 2x oversampling from 44.1 kHz, processing happens at 88.2 kHz.
- Process: Apply the nonlinear algorithm at the higher rate. The Nyquist ceiling is now 44.1 kHz instead of 22.05 kHz.
- Lowpass filter: Cut everything above 22.05 kHz (the original Nyquist). This removes any remaining aliases before they can fold into the audible range during step 4.
- Downsample: Return to the original 44.1 kHz sample rate.
Returning to the 8 kHz example at 2x oversampling:
- 24 kHz: below the new Nyquist of 44.1 kHz, no aliasing
- 32 kHz: below 44.1 kHz, no aliasing
- 40 kHz: below 44.1 kHz, no aliasing
- 48 kHz: above 44.1 kHz, alias at 88.2 - 48 = 40.2 kHz, but 40.2 kHz is well above the 22.05 kHz lowpass cutoff applied during decimation and gets removed
With 2x oversampling, the first problematic aliasing artifact from an 8 kHz input shifts from 20 kHz (audible) into territory that the decimation filter discards. For most musical material, 2x eliminates the perceptible harshness. Aggressive distortion on signals with significant high-frequency content, a bowed string run through heavy asymmetric clipping for instance, may benefit from 4x.
FabFilter Saturn 2 exposes this tradeoff directly in its interface with labeled oversampling tiers, from a light mode for CPU-sensitive situations to a maximum-quality mode intended for final mixdowns.
The Filter Tradeoff: Linear Phase vs. Minimum Phase
The decimation lowpass filter that strips harmonics above 22.05 kHz is not free. How it is designed determines the sonic and latency character of the oversampled processor.
A linear phase filter has constant group delay across all frequencies, meaning all spectral components are delayed by the same amount. The impulse response is symmetric, introducing pre-ringing before transients. For offline processing this is generally the best choice because phase relationships are preserved exactly.
A minimum phase filter introduces no pre-ringing. All ringing occurs after the event. Group delay is not constant, meaning different frequencies arrive at slightly different times, but for most saturation and distortion processors this is perceptually irrelevant. The latency is minimal, which matters in real-time contexts and when comparing dry versus processed signals in a parallel routing.
Most plugin implementations use minimum phase decimation filters in their real-time oversampling modes.
CPU Cost and What It Actually Scales With
CPU usage from oversampling scales linearly with the oversampling ratio for the processing stages involved. 2x oversampling of a single plugin costs roughly twice the CPU of that same plugin at 1x. This is not the full picture, though: the anti-aliasing lowpass filter itself carries a cost that depends on its steepness. A sharper transition band requires a longer filter, which means more computation.
Critically, oversampling only needs to be applied to the nonlinear stages of a signal chain. A processor that internally upsamples just its waveshaping core, while running the input/output gain stages and metering at native rate, can achieve aliasing reduction at a fraction of the CPU cost of naively oversampling everything.
The ADAA Alternative
Oversampling is the industry standard solution, but it is not the only one. Antiderivative Antialiasing, referred to as ADAA in DSP literature, takes a different approach: instead of processing at a higher rate, it reformulates the nonlinear function mathematically so that it does not generate aliases in the first place.
Research from Aalto University and published AES papers by Välimäki and colleagues has shown that ADAA can achieve significant alias reduction at 1x sample rate for a specific class of memoryless nonlinearities. The limitation is that it only works for functions that have a tractable antiderivative and no memory between samples. Hard clipping works well. Feedback-based analog circuit models generally do not.
Some newer plugin architectures use ADAA for their waveshaping cores, reducing CPU load compared to equivalent oversampling ratios. It is an active area of research, particularly in neural network-based amp simulations where standard oversampling is expensive.
When Not to Use Oversampling
Linear time-invariant processors do not generate new frequencies. An EQ, a delay line, a clean gain stage, or a reverb algorithm built from linear filters will produce no aliasing regardless of sample rate. Enabling oversampling on these processors adds CPU cost with zero audible benefit.
The practical rule: enable oversampling on anything that multiplies, clips, folds, or otherwise performs a nonlinear operation on the signal. Leave it off everywhere else.
At 4x oversampling, the aliases from most real-world audio material are attenuated below perceptibility for the entire harmonic series up to the 10th or 12th harmonic. For mixing and mastering at 44.1 or 48 kHz, this is more than sufficient. Processing at 88.2 kHz or 96 kHz natively already shifts the aliasing problem up by an octave, at which point even 2x oversampling eliminates every alias above the threshold of hearing.