Multirate Signal Processing

Changing the sampling rate of a digital signal is a fundamental operation in modern engineering systems, allowing different parts of a system to operate at their most efficient speed. Whether you're converting between digital audio standards, reducing computational load for a sensor, or preparing data for transmission, multirate signal processing provides the theoretical and practical toolkit, covering core operations such as decimation and interpolation, the clever filter structures that make them efficient, and their indispensable role in real-world applications.

Review: The Sampling Theorem and Aliasing

To understand why changing a sample rate isn't as simple as throwing away or adding points, you must recall the Nyquist-Shannon sampling theorem. It states that a continuous-time signal can be perfectly reconstructed from its samples if it is sampled at a rate $f_s$ greater than twice its highest frequency component $f_{max}$. The frequency $f_s/2$ is called the Nyquist frequency.

If a signal contains energy above the Nyquist frequency before sampling, or if such frequencies are created by a processing step like downsampling, aliasing occurs. Aliasing is the phenomenon where these high frequencies masquerade as lower, incorrect frequencies within the baseband, irreparably corrupting the signal. Preventing aliasing is the primary driver behind the careful procedures in multirate processing.

Core Operation 1: Decimation (Downsampling)

Decimation is the process of reducing the sampling rate of a signal by an integer factor $M$. The naive approach—simply keeping every $M$-th sample—is called downsampling and is denoted by the operator $\downarrow M$. However, this direct approach is dangerous. Downsampling by $M$ reduces the new Nyquist frequency to $f_s^{\prime }/2=f_s/(2M)$. Any signal component in the original signal above this new, lower Nyquist frequency will alias into the baseband.

Therefore, decimation is always a two-step process:

1. Anti-aliasing Filtering: First, the original signal is passed through a lowpass filter called a decimation filter. This filter is designed with a cutoff frequency at or below $f_s/(2M)$. Its job is to aggressively attenuate all frequency components above the new Nyquist frequency to a negligible level.
2. Downsampling: The filtered signal, now bandlimited appropriately, is passed through the $\downarrow M$ operator. Every $M$-th sample is retained, and the others are discarded, producing the output at the lower rate $f_s^{\prime }=f_s/M$.

The combined operation is efficient because the filter operates at the high input rate $f_s$, but we only compute the output samples that will be kept after downsampling. This leads directly to the efficiency of polyphase structures, discussed later.

Core Operation 2: Interpolation (Upsampling)

Interpolation is the process of increasing the sampling rate of a signal by an integer factor $L$. The complementary naive operation is upsampling, denoted by $\uparrow L$, which inserts $L-1$ zeros between each original sample. This increases the sample rate to $f_s^{\prime }=L{f}_s$.

While upsampling alone increases the sampling rate, it creates spectral copies (or "images") of the original signal's spectrum at multiples of the original sampling frequency. These are artifacts that must be removed. Therefore, interpolation is also a two-step process:

1. Upsampling: Insert $L-1$ zeros between each original sample. This creates the higher-rate sequence but with the unwanted spectral images.
2. Anti-imaging Filtering: The upsampled signal is passed through a lowpass filter called an interpolation filter. This filter has a cutoff frequency of $f_s/2$ (the original Nyquist frequency) and is designed to smooth the zero-stuffed signal, removing the spectral images and effectively calculating the interpolated values for the inserted points.

Polyphase Filter Structures: The Engine of Efficiency

A direct implementation of the decimation filter is wasteful: it computes all filtered samples at the high rate $f_s$ only to discard $M-1$ out of every $M$ outputs. Similarly, an interpolation filter multiplies many zero-valued samples from the upsampler. Polyphase filter structures eliminate this redundancy by reorganizing the filtering operations to run at the lower of the two involved sampling rates.

The key idea is to decompose the original lowpass filter impulse response $h[n]$ into $M$ (for decimation) or $L$ (for interpolation) distinct sub-filters, called polyphase components. Each component is a downsampled version of a delayed $h[n]$.

• For Decimation: The input data is demultiplexed into $M$ parallel streams. Each stream is filtered by its corresponding polyphase component, but crucially, this filtering now happens at the output (low) rate. The results are summed to produce the final decimated output. This means all multiplications and additions are performed at the lower rate $f_s/M$, saving substantial computation.
• For Interpolation: The process is essentially the inverse. The input at the low rate is fed into $L$ parallel polyphase filters. Their outputs are multiplexed (interleaved) to create the high-rate output sequence. Again, all filtering work is done at the lower input rate, even though the final output is at the high rate $L{f}_s$.

Polyphase implementations are the standard in practical systems due to this dramatic improvement in computational efficiency, often by a factor equal to the rate change.

Common Pitfalls

1. Skipping or Designing a Poor Anti-aliasing Filter: The most critical error is downsampling without proper anti-aliasing filtration. Even mild signal content above the new Nyquist frequency will cause aliasing distortion that cannot be removed later. Always verify your filter's stopband attenuation is sufficient for your application's requirements.
2. Confusing Rate Change Order in Cascade: When performing multiple rate changes (e.g., a conversion by a non-integer factor like $4/3$), the order of filtering operations is vital. A standard efficient approach is to perform interpolation first (to raise the rate to a least common multiple), apply a single filter at the high rate, and then decimate. Placing a decimation step before necessary filtering will cause aliasing.
3. Ignoring Computational Efficiency: Implementing decimation or interpolation using a straightforward filter followed by a down/up-sampler in a software loop or naive hardware design wastes immense computational resources. Always consider a polyphase or other multirate-optimized filter structure for real-world implementations.
4. Misunderstanding the "Images" in Interpolation: The spectral images created by upsampling are not aliasing; they are exact copies. However, they are almost always undesirable. The anti-imaging filter is sometimes called a "smoothing filter" because its time-domain role is to fill in the zeros with interpolated values, which in the frequency domain corresponds to removing these images.

Summary

• Decimation reduces sampling rate by first lowpass filtering (anti-aliasing) to prevent high frequencies from aliasing, then discarding samples. It is denoted as filtering followed by $\downarrow M$.
• Interpolation increases sampling rate by first inserting zeros (upsampling $\uparrow L$), then lowpass filtering (anti-imaging) to remove spectral copies and smooth the signal.
• Polyphase filter structures reorganize the filtering operations to run at the lower data rate in the system, providing massive computational savings and are the standard for efficient implementation.
• The core principle governing all operations is strict adherence to the Nyquist criterion to avoid aliasing during downsampling and to remove imaging artifacts during upsampling.
• These multirate techniques are essential in audio sample-rate conversion, digital communication transceivers (for matched filtering and channelization), and sensor/control systems where processing bandwidth must be matched to signal bandwidth.