subsampling

Subsampling is the process of selecting a subset of samples from a larger dataset or signal, typically to reduce size, lower computational cost, or alter the sampling rate. Subsampling can be applied across disciplines, including signal processing, statistics, machine learning, and imaging. The chosen subset should be representative to preserve essential information.

In signal processing, subsampling is often called downsampling or decimation. The sampling rate is reduced by keeping every M-th sample. Before decimation, a low-pass filter is usually applied to prevent aliasing. The new sampling rate should be at least twice the highest frequency present in the signal.

In statistics and machine learning, subsampling refers to selecting a subset of observations. This is common for creating training and testing splits, reducing computation, or addressing class imbalance by undersampling the majority class. Subsampling can increase variance and introduce bias if the subset is not representative; methods such as random, stratified, or systematic subsampling are used to improve representativeness. Bootstrapping is a related but distinct resampling technique that samples with replacement.

In imaging and video, subsampling can refer to chroma subsampling in color spaces (for example, 4:2:0), where color resolution is reduced relative to luminance to save bandwidth or storage. Subsampling also occurs when creating lower-resolution images by sampling every k-th pixel.

Limitations include loss of information, potential bias, and the need to consider the temporal or spatial structure of the data. Best practices involve choosing an appropriate factor, preserving representativeness, and validating results on the full dataset when possible.