Also known as: oversampled conversion
Oversampling means sampling a signal at a rate substantially higher than the minimum the Nyquist theorem requires — often many times higher.1 The extra samples are not wasted: because a converter’s total quantization noise is fixed in power but now spread across a much wider band, only a fraction of it falls in the band of interest. Filtering off the rest and decimating back down recovers resolution, relaxes the analog anti-alias filter, and cleans up the signal.
How it works
The oversampling ratio (OSR) is the actual sample rate divided by twice the signal bandwidth. Quantization noise power is essentially constant regardless of rate, so sampling faster distributes that power over a proportionally wider spectrum — the power spectral density of the noise drops. A digital low-pass filter then keeps only the signal band and rejects the noise outside it. Each doubling of the sample rate removes half the in-band noise, worth about 3 dB, or half a bit of extra effective resolution — this is the processing gain of oversampling. After filtering, decimation throws away the now redundant samples to return to an efficient rate.
Noise shaping does far better. A delta-sigma (Σ-Δ) modulator feeds back the quantization error so that most of the noise is pushed up to high frequencies, away from the signal band, before the decimation filter removes it. With shaping, each doubling of OSR can buy far more than 3 dB — this is how a 1-bit Σ-Δ converter achieves 16 or more effective bits at audio bandwidths.
In practice
Beyond resolution, oversampling relaxes the analog anti-alias filter. At the Nyquist rate the filter must transition from passband to full rejection within a razor-thin guard band, demanding a steep, expensive analog design. Oversample, and the first alias sits far away, so a gentle analog filter suffices and the sharp cutoff is done digitally where it is cheap and precise. Oversampling also pairs with dither: the added noise keeps the quantization error smooth so the processing gain is genuine noise reduction rather than smeared distortion.
The cost is throughput — more samples per second to move and process — which is why systems oversample at the ADC and then decimate to a manageable working rate.
Relevance to SDR
Oversampling is pervasive in SDR. Converters run far above the wanted channel bandwidth, and the receiver’s digital down-converter and decimating filters recover both processing gain and a clean channel at a lower rate. It is also why capturing at a higher sample rate than a single channel needs can help: the wider capture spreads quantization noise, and the channelising filter reclaims dynamic range for the channel you keep. The same decimation chains built on CIC and half-band filters that step a wide capture down to a channel rate are the practical machinery of oversampling.
GopherTrunk’s DSP chain oversamples in exactly this sense: it takes a wide IQ capture and decimates each channel down to the per-protocol symbol rate, so the filtering that isolates a control channel also delivers the processing gain that oversampling promises.
Sources
-
Oversampling — Wikipedia, on sampling above Nyquist to spread quantization noise, relax anti-aliasing, and gain resolution by decimation. ↩