4. Multirate Signal Processing and Sampling-Rate Conversion

Introduction and Context

Digital signal processing traditionally assumes that a signal is processed at a fixed sampling frequency. Many practical systems, however, must exchange information between subsystems operating at different rates. A communication receiver may convert a radio-frequency-derived sample stream to a lower baseband rate; an audio interface may translate between 44.1 kHz and 48 kHz; an image system may change spatial resolution; and a software-defined radio may use several rates as signals pass through filtering, synchronization, detection, and decoding stages.

Multirate signal processing provides the mathematical and architectural tools for such systems. Its central operations are decimation, which reduces the sampling rate, and interpolation, which increases it. Combined appropriately, these operations realize rational sampling-rate conversion. The subject is not merely a matter of inserting or deleting samples. Rate conversion changes the discrete-time frequency axis, and unless spectral overlap is carefully controlled, it produces aliasing or imaging distortion.

The principal engineering challenge is therefore to obtain the required rate while preserving the desired signal spectrum with acceptable computational cost, delay, numerical accuracy, and implementation complexity. This chapter develops the theory of sampling-rate conversion, derives efficient polyphase structures, explains noble identities, and examines multistage designs used in advanced digital systems.

Learning Objectives

After studying this chapter, the reader should be able to:

1. distinguish between decimation, interpolation, and rational rate conversion;

2. explain how frequency spectra transform under sampling-rate changes;

3. design anti-aliasing and anti-imaging filters from transition-band requirements;

4. derive and apply polyphase decompositions;

5. use noble identities to move filters across rate-changing blocks;

6. compare direct, polyphase, and multistage implementations;

7. evaluate computational load, delay, aliasing, and implementation trade-offs; and

8. select an appropriate sampling-rate conversion architecture for a practical system.

The Sampling-Rate Conversion Problem

Let x[n] be a discrete-time signal sampled at frequency F_s. Its discrete-time Fourier transform (DTFT) is periodic in angular frequency with period 2π. A change in sampling frequency changes the physical frequency represented by a given normalized frequency. Consequently, sampling-rate conversion must be analyzed in both normalized frequency, measured in radians per sample, and physical frequency, measured in hertz.

A useful design principle is that the information-bearing spectrum should occupy only the frequency interval that remains unambiguous after conversion. For a real-valued signal with highest significant frequency F_B, the original sampling rate must satisfy F_s > 2F_B. During conversion, the relevant Nyquist frequency becomes F_s'/2, where F_s' is the new rate. If F_s' is lower than the original rate, components above F_s'/2 cannot remain distinct and must be removed before downsampling.

Two types of spectral error result from rate changes:

• Aliasing: spectral components overlap after a rate reduction and cannot be separated by later processing.
• Imaging: spectral replicas appear in the spectrum after a rate increase because the newly created samples do not yet define a smooth band-limited sequence.

Both effects are consequences of spectral periodicity, but their locations in a processing chain differ. Aliasing is prevented before decimation; imaging is suppressed after interpolation.

Decimation: Reducing the Sampling Rate

Decimation by an integer factor M consists conceptually of low-pass filtering followed by downsampling by M. If v[n] is the filtered sequence, the output is

y[n] = v[nM].

The downsampler retains one sample out of every M samples, so the output sampling frequency is F_s/M. The operation is sometimes called downsampling, although strictly speaking decimation includes the anti-aliasing filter as well.

The frequency-domain relation for downsampling is

Y(e^{jω}) = (1/M) Σ_{k=0}^{M-1} V(e^{j(ω+2πk)/M}).

Thus, the output spectrum is a scaled sum of M shifted and compressed copies of the input spectrum. If these copies overlap within the output Nyquist interval, aliasing occurs. For an ideal decimator, the preceding filter should satisfy

V(e^{jω}) = 0 for |ω| ≥ π/M.

In physical frequency, this means that the input signal must be limited to below F_s/(2M), which equals half the output sampling rate. A practical anti-aliasing filter has a passband edge below this limit and a stopband beginning no later than the frequency that would alias into the desired band.

Example: Integer Decimation

Suppose a 96 kHz sampled signal is to be decimated by M = 4. The output rate is 24 kHz and its Nyquist frequency is 12 kHz. Frequencies above 12 kHz must be attenuated before samples are discarded. If the required signal band extends to 10 kHz, the filter passband may end at 10 kHz and its stopband must begin at 12 kHz, subject to the specified ripple and attenuation. A filter designed without this precaution would permit components between 12 and 48 kHz to fold into the baseband.

Interpolation: Increasing the Sampling Rate

Interpolation by an integer factor L consists conceptually of upsampling followed by low-pass filtering. Upsampling creates the intermediate sequence

v[n] = x[n/L], when n is a multiple of L,

and v[n] = 0 otherwise. The sampling rate becomes LF_s, but the inserted zeros do not represent estimates of the signal at the new time instants. They create periodic spectral images.

The DTFT of the zero-inserted signal is

V(e^{jω}) = X(e^{jLω}).

Therefore, the original spectrum is compressed by L in normalized frequency, producing L repeated images over the interval −π ≤ ω < π. A reconstruction or anti-imaging low-pass filter removes these images. The ideal interpolation filter has gain L in the retained band, compensating for the 1/L scaling associated with the spectral replicas.

Interpolation by L is therefore not simply sample insertion. Its meaningful operation is:

1. increase the sampling rate by inserting zeros;

2. low-pass filter the resulting sequence; and

3. preserve the original spectral content while eliminating images.

Example: Integer Interpolation

An 8 kHz speech signal is interpolated by L = 3 to produce a 24 kHz sequence. The original speech spectrum occupies a narrow band relative to the new Nyquist frequency of 12 kHz. Zero insertion creates two additional spectral images. A low-pass filter with passband covering the speech band and stopband beginning before the first image removes those replicas.

Rational Sampling-Rate Conversion

Many practical conversions require a noninteger ratio. Let the desired output rate be related to the input rate by

F_out/F_in = L/M,

where L and M are positive integers, normally chosen relatively prime. The standard structure first interpolates by L and then decimates by M. The intermediate rate is LF_in, and the final rate is

F_out = (L/M)F_in.

A single low-pass filter between these operations can perform both anti-imaging and anti-aliasing functions. Its cutoff must be chosen in the intermediate-rate domain. In normalized frequency relative to the intermediate sampling rate, the passband edge is limited by

ω_p < min(π/L, π/M)

when the input and output spectra are specified in normalized terms. In physical frequency, the retained bandwidth must fit below both the original and final Nyquist constraints.

Example: 44.1 kHz to 48 kHz

The ratio is

48/44.1 = 160/147.

Thus, a direct rational converter interpolates by L = 160 and decimates by M = 147. Although conceptually straightforward, this would create a high intermediate rate of 7.056 MHz. A direct implementation would perform many unnecessary multiplications, especially because most interpolated samples are eventually discarded. Polyphase and multistage methods make this conversion practical.

Anti-Aliasing and Anti-Imaging Filter Design

Filter design begins with four quantities: passband edge, stopband edge, passband ripple, and stopband attenuation. For a decimator, the stopband must begin before frequencies that fold into the retained band. For an interpolator, the stopband must suppress the first spectral image and all subsequent images within the digital frequency interval.

The transition width determines filter order. A narrower transition band generally requires more coefficients. For an FIR filter, a common approximate relationship is

N ≈ (A_s − 8)/(2.285 Δω),

where N is the order, A_s is the required stopband attenuation in decibels, and Δω is the transition width in radians per sample. This relation is useful for preliminary estimates rather than final design.

FIR filters are often preferred in sampling-rate conversion because they can provide exact linear phase, stable finite-duration responses, and efficient polyphase implementations. IIR filters may achieve lower order, but their phase response, coefficient sensitivity, and state handling across rate changes require additional care. In high-performance communication and audio systems, the choice depends on latency, phase requirements, attenuation, power consumption, and hardware resources.

Polyphase Decomposition: The Main Efficiency Principle

The direct rational converter forms the high-rate sequence, filters it, and then retains only selected output samples. This is wasteful: most intermediate samples are either zeros or discarded. Polyphase decomposition reorganizes the filter so that only samples contributing to actual outputs are computed.

For an FIR filter

h[n], 0 ≤ n < N,

the L polyphase components with respect to interpolation by L are

E_k[r] = h[rL + k], k = 0, 1, ..., L−1.

The filter can be written as

H(z) = Σ_{k=0}^{L−1} z^{−k} E_k(z^L).

In an interpolator, each output phase uses one of these subfilters. Because zero-valued input positions do not contribute to ordinary convolution, the implementation avoids multiplying by inserted zeros.

For decimation by M, the filter can instead be partitioned into M phases:

E_k[r] = h[rM + k], k = 0, 1, ..., M−1.

The output samples are computed directly from the relevant phase, avoiding the generation of samples that the decimator would immediately discard. A rational converter may use a bank of phases associated with the interpolation factor and select phases according to the decimation schedule.

Time-Index Interpretation

The output sample at index n corresponds to an input-time position determined by the ratio L/M. A phase accumulator tracks this position. At each output instant, the accumulator identifies the appropriate fractional-delay phase, while the filter bank supplies the corresponding coefficient set. This interpretation is especially useful in software implementations, asynchronous sample-rate conversion, and hardware architectures with coefficient memory.

Polyphase processing reduces the approximate number of multiplications per output sample from a cost associated with the high intermediate rate to a cost proportional to the number of coefficients actually used for that output. It also supports coefficient symmetry, SIMD execution, parallel hardware, and controlled memory access.

Noble Identities and Structural Transformations

Noble identities are algebraic rules that allow filtering and rate changing to be reordered without changing the overall input-output relation, provided the filter transfer function is correspondingly transformed.

For decimation by M, filtering before downsampling can be represented by moving the downsampler before a modified filter whose coefficients are expressed in the appropriate polyphase form. Informally, a filter operating at the high rate need not process every high-rate sample if only every Mth output is needed.

For interpolation by L, an interpolator can be moved across a filter by replacing the filter with a version in which the delay variable is appropriately expanded. This permits filtering to occur at the lower input rate rather than at the higher output rate when the structure allows it.

The noble identities are more than symbolic manipulations. They provide a systematic method for transforming an obvious but inefficient block diagram into an efficient one. A designer should first establish the required overall transfer function and rate relationship, then move rate-changing blocks across filters, decompose the filters into phases, and verify that the resulting structure preserves gain, delay, and frequency response.

Multistage Conversion Architectures

A large conversion factor is rarely best implemented as one stage. If the ratio can be factored into smaller integers, the conversion may be divided into stages. For example, decimation by 60 can be realized as successive factors 3, 4, and 5. Each stage then requires a filter designed for its local rate and transition band.

Multistage design offers several benefits:

• filters operate at lower rates whenever possible;
• transition bands may be wider in normalized frequency;
• coefficient counts and arithmetic rates can be reduced;
• intermediate word lengths and hardware resources can be allocated stage by stage; and
• different stages can use different filter types.

The best factorization is not determined by arithmetic alone. A stage with a very narrow transition band may dominate the total cost. The designer must consider cumulative passband ripple, stopband attenuation, delay, coefficient quantization, and the location of each stage relative to the signal bandwidth.

Half-band filters are especially valuable for factors of two. Ideally, approximately half of their coefficients are zero, and many of the remaining coefficients can be paired through symmetry. Cascaded integrator-comb (CIC) filters are attractive for large integer rate changes in hardware because they use only adders and delays, although they have substantial passband droop and require compensation when flat amplitude response is important.

A rational conversion can also be factored into several rational stages. For example, 160/147 may be implemented using factors selected from 2, 4, 5, 7, and related combinations. The exact architecture depends on whether the system prioritizes minimum power, minimum latency, high stopband attenuation, or simple clocking.

An Applied Design Scenario: Software-Defined Radio Receiver

Background

Consider an illustrative software-defined radio receiver whose front end produces complex samples at 61.44 MS/s. A downstream demodulator requires 3.84 MS/s. The overall decimation factor is 16.

Situation and Problem

The front-end bandwidth is wider than the demodulator bandwidth, but the receiver must prevent out-of-band noise and adjacent-channel energy from aliasing into the desired channel. A single filter followed by decimation by 16 is possible, but its transition band may be demanding and its high-rate computation expensive.

Analysis

A factorization such as 2 × 2 × 2 × 2 permits four stages. The first stages operate at high rates and must remove energy that could fold during the next decimation. Later stages operate at progressively lower rates and can use filters tailored to the narrower signal bandwidth. If the desired channel occupies a small fraction of the 3.84 MS/s Nyquist interval, the final stage can provide strong adjacent-channel rejection.

Approach

The receiver uses a cascade of polyphase half-band FIR filters for the factor-of-two stages, followed by a final sharper low-pass filter if required by the channel specification. Each stage is designed with a passband edge and stopband edge referenced to its own input rate. The total allowable passband ripple is distributed across the stages, and stopband attenuation is selected so that aliased contributions remain below the receiver noise and interference budget.

Fixed-point simulation is performed using representative multichannel signals, blockers, and quantization noise. Coefficients are quantized, accumulator widths are selected to prevent overflow, and the group delay is included in the synchronization budget.

Outcome

The multistage polyphase structure reduces the number of multiplications compared with direct high-rate filtering and makes the clocking requirements manageable. The receiver preserves the desired channel while suppressing energy that would otherwise alias after each rate reduction.

Learning Points

The important design lesson is that rate conversion is a system-level filtering problem. The conversion factor alone does not determine the architecture. Signal bandwidth, interference environment, transition width, hardware resources, latency, and numerical precision must be considered together.

Implementation Considerations and Common Errors

A frequent mistake is to downsample first and filter afterward. Once aliasing has occurred, no subsequent linear filter can reconstruct the original spectral components. Similarly, zero insertion without anti-imaging filtering produces a sequence containing spectral images rather than a properly interpolated signal.

Other common errors include confusing normalized and physical frequency, forgetting the gain factor associated with interpolation, using an incorrect phase in a polyphase bank, and overlooking filter delay when aligning parallel paths. In rational conversion, the numerator and denominator must be assigned correctly: L increases the rate and M decreases it.

Finite-precision implementation introduces coefficient quantization, round-off noise, overflow, and limit-cycle concerns in recursive filters. FIR structures are generally easier to analyze, but long filters may require substantial memory and multiply-accumulate throughput. Saturation should be specified deliberately rather than allowed to occur unpredictably.

Clock-domain boundaries require additional engineering attention. In a synchronous converter, the ratio may be exact and periodic. In an asynchronous sample-rate converter, the ratio varies or is not rational over the observation interval. Such systems use variable fractional-delay filters, timing estimators, and phase accumulators. The filter must then maintain adequate attenuation over all possible phases.

Professional and Contemporary Applications

Multirate methods are fundamental in digital communication channelizers, software-defined radios, audio sample-rate converters, hearing aids, digital instrumentation, sensor fusion, ultrasound, radar, image pyramids, and oversampled analog-to-digital and digital-to-analog converters. Filter banks and multirate analysis-synthesis systems extend the same principles to subband coding and wavelet-like representations.

Modern implementations increasingly exploit programmable logic, digital signal processors, graphics processors, and vector instruction sets. Automatic design tools can optimize stage factorization, coefficient quantization, and polyphase scheduling under constraints on power, latency, and memory. Machine-learning methods may assist in system-level optimization, but they do not remove the requirement for mathematically verified aliasing and imaging limits. Safety-critical and communication systems must retain auditable frequency-response specifications and reproducible verification procedures.

Design Best Practices

A reliable sampling-rate converter is developed through the following reasoning sequence:

1. specify input and output rates and reduce the ratio to relatively prime integers;

2. define the desired passband, transition band, ripple, and stopband attenuation in physical frequency;

3. identify whether each stage is vulnerable to aliasing or imaging;

4. choose direct, polyphase, or multistage realization according to computational and latency constraints;

5. apply noble identities and exploit coefficient symmetry or sparsity;

6. simulate frequency response, impulse response, tones near band edges, blockers, and quantization effects;

7. verify gain, phase, delay, overflow margins, and cumulative attenuation; and

8. test the implementation with hardware-rate clocks and worst-case signals.

These steps prevent a common design failure: optimizing arithmetic before establishing the correct spectral specification.

Chapter Summary

Multirate signal processing enables controlled movement between sampling frequencies. Decimation reduces the rate and requires anti-aliasing filtering; interpolation increases the rate and requires anti-imaging filtering. Rational conversion by L/M combines interpolation by L with decimation by M, usually through a shared low-pass filter. The spectral relations explain why aliasing and imaging occur and determine the filter constraints.

Polyphase decomposition eliminates computations involving inserted zeros and discarded intermediate samples. Noble identities provide the algebraic basis for moving filters across rate-changing operations and obtaining efficient structures. For large factors, multistage architectures, half-band filters, and CIC filters can substantially reduce cost, but their transition bands, ripple, attenuation, delay, and numerical effects must be evaluated together. In professional systems, sampling-rate conversion is a complete architectural problem involving spectral integrity, implementation efficiency, synchronization, and verification.

Key Terms

Decimation: Reduction of a sampling rate, normally by filtering followed by downsampling.

Downsampling: Selection of every Mth sample from a discrete-time sequence.

Interpolation: Increase of a sampling rate through zero insertion and anti-imaging filtering.

Upsampling: Insertion of zeros between consecutive samples.

Rational rate conversion: Conversion whose output-to-input rate ratio is L/M.

Aliasing: Irrecoverable spectral overlap caused by insufficient pre-decimation filtering.

Imaging: Spectral replicas produced by upsampling before anti-imaging filtering.

Anti-aliasing filter: Low-pass filter applied before decimation.

Anti-imaging filter: Low-pass filter applied after interpolation.

Polyphase filter: A filter decomposed into subfilters corresponding to different sample phases.

Noble identity: An equivalence that permits a filter and a rate changer to be reordered.

Passband: Frequency range intended to pass with specified amplitude accuracy.

Stopband: Frequency range in which attenuation is required.

Transition band: Frequency interval between passband and stopband.

Half-band filter: Low-pass filter designed around half the Nyquist frequency, often with many zero coefficients.

CIC filter: Multistage integrator-comb filter using additions and delays for efficient rate change.

Fractional-delay filter: Filter that estimates a signal at a noninteger time position.

Phase accumulator: State variable used to track the required phase in rational or asynchronous conversion.

Passband droop: Frequency-dependent attenuation within the intended passband.

Group delay: Frequency-dependent or constant delay associated with a filter's phase response.

Review Questions

Short Answer Questions

1. Why must an anti-aliasing filter precede an integer decimator?

2. What spectral effect is caused by zero insertion during interpolation?

3. State the relationship between input and output rates for rational conversion.

4. What computational waste is removed by polyphase implementation?

5. Why are half-band filters useful in cascaded factor-of-two conversion?

Long Answer Questions

1. Derive the spectrum of a sequence downsampled by M and explain the origin of aliasing.

2. Explain interpolation by L, including zero insertion, spectral images, and filter gain.

3. Describe the construction and operation of a polyphase rational-rate converter.

4. Explain how noble identities lead to efficient multirate block diagrams.

5. Discuss the design trade-offs between direct, polyphase, and multistage conversion architectures.

Analytical / Application Questions

1. Design the conceptual filtering and rate-changing sequence required to convert 96 kHz data to 24 kHz while preserving a 10 kHz signal band.

2. A system converts 44.1 kHz audio to 48 kHz. Determine suitable interpolation and decimation factors and explain why direct implementation is inefficient.

3. Propose a multistage architecture for decimation by 16 in a software-defined radio, including filter allocation, aliasing control, and implementation considerations.

22. Multiple Choice Questions with Answers

Question 1

What is the principal purpose of filtering before decimation by M?

A. To create spectral images

B. To prevent spectral overlap after rate reduction

C. To increase the sampling rate

D. To introduce zeros between samples

Correct Answer: B. To prevent spectral overlap after rate reduction

Question 2

If a signal sampled at 96 kHz is decimated by 4, what is the output sampling rate?

A. 12 kHz

B. 24 kHz

C. 48 kHz

D. 384 kHz

Correct Answer: B. 24 kHz

Question 3

What happens to the spectrum immediately after upsampling by L through zero insertion?

A. It is compressed and repeated, producing images

B. It is eliminated completely

C. It is expanded without replicas

D. It becomes nonperiodic

Correct Answer: A. It is compressed and repeated, producing images

Question 4

For conversion from 44.1 kHz to 48 kHz, which relatively prime factors implement the rate ratio?

A. 147/160

B. 48/44.1

C. 160/147

D. 44.1/48

Correct Answer: C. 160/147

Question 5

What is the main computational advantage of a polyphase interpolator?

A. It removes the need for any low-pass filter

B. It computes products involving inserted zeros

C. It avoids unnecessary operations associated with zero samples

D. It always provides nonlinear phase

Correct Answer: C. It avoids unnecessary operations associated with zero samples

Question 6

Which phenomenon cannot be corrected by a later linear filter once it has occurred?

A. Passband ripple

B. Aliasing

C. Group delay

D. Imaging before filtering

Correct Answer: B. Aliasing

Question 7

A converter has F_out/F_in = L/M. Which operation increases the intermediate sampling rate?

A. Decimation by M

B. Interpolation by L

C. Filtering by M

D. Phase accumulation by M

Correct Answer: B. Interpolation by L

Question 8

Why are multistage architectures often preferred for a large rate-change factor?

A. They eliminate all filter specifications

B. They allow filtering to be distributed across lower-rate stages

C. They guarantee zero delay

D. They require no coefficient quantization

Correct Answer: B. They allow filtering to be distributed across lower-rate stages

Question 9

What is a significant limitation of a CIC filter?

A. It requires floating-point multiplication for every sample

B. It cannot change a sampling rate

C. It can exhibit passband droop requiring compensation

D. It produces no frequency response

Correct Answer: C. It can exhibit passband droop requiring compensation

Question 10

In an asynchronous sample-rate converter, what does a fractional-delay filter provide?

A. An estimate at a noninteger time position

B. A permanent increase in signal bandwidth

C. Removal of all quantization noise

D. A replacement for the sampling clock

Correct Answer: A. An estimate at a noninteger time position