Electronics Guide

Sample Rate and the Nyquist Theorem

The Nyquist-Shannon sampling theorem establishes that a bandlimited signal can be perfectly reconstructed from its samples if the sampling rate exceeds twice the signal's highest frequency component. This critical frequency, the Nyquist frequency, equals half the sampling rate. Any frequency content above the Nyquist frequency cannot be distinguished from lower frequencies, a phenomenon called aliasing that irreversibly corrupts the signal.

When a signal's bandwidth is significantly less than the Nyquist frequency, it is oversampled relative to its information content. This excess sample rate represents wasted computational resources when processing the signal. Multirate techniques exploit this observation by reducing the sample rate after filtering removes high-frequency components, computing subsequent operations at the minimum rate necessary to preserve signal fidelity.

Conversely, signals sometimes must be represented at higher sample rates than their bandwidth strictly requires. Interfacing with digital-to-analog converters, combining signals sampled at different rates, or preparing signals for further processing may all require increased sample rates. Interpolation increases rates by computing new samples between existing ones, effectively filling in the signal waveform at finer time resolution.

Spectral Effects of Rate Changes

Downsampling by a factor M, which retains every Mth sample while discarding the rest, stretches the signal's spectrum by factor M along the normalized frequency axis. Content that occupied the band from zero to the new Nyquist frequency, one Mth of the original Nyquist frequency, expands to fill the entire baseband at the reduced rate. Any energy that lay above the new Nyquist frequency folds back, or aliases, into the baseband and corrupts the signal. Anti-aliasing filters must remove this high-frequency content before downsampling.

Upsampling by a factor L, which inserts L − 1 zeros between each original sample, compresses the spectrum by factor L and thereby brings L − 1 spectral images of the baseband into view, centered at multiples of the original sampling frequency. These images are simply the periodic replicas of the original spectrum, always present in a sampled signal, that now fall within the wider Nyquist band created by the higher rate. Interpolation filtering removes the images, leaving only the baseband spectrum while filling in the time-domain signal values at the new sample positions.

The frequency scaling associated with rate changes requires careful attention when designing multirate systems. Filter specifications must account for the relationship between continuous frequency and normalized digital frequency at each sample rate in the system. A filter specified in terms of continuous frequency (hertz) will have different normalized cutoff frequencies at different sample rates, a fact that must be considered when cascading rate changes with filtering operations.

Basic Building Blocks

The downsampler, symbolized by a down-arrow with the decimation factor M, reduces sample rate by keeping every Mth sample. Mathematically, if the input sequence is x[n], the output is y[m] = x[mM]. This operation is memoryless in the sense that each output depends only on one input, but it is not reversible because information is discarded.

The upsampler, symbolized by an up-arrow with the interpolation factor L, increases sample rate by inserting L − 1 zeros between each input sample. If the input is x[n], the output is y[m] = x[m/L] when m is a multiple of L, and zero otherwise. Like downsampling, upsampling is memoryless, but unlike downsampling, no information is lost because the original samples can be recovered by downsampling the result.

Filters in multirate systems serve different purposes depending on their location relative to rate changes. Anti-aliasing filters before downsamplers remove frequency content that would alias. Interpolation filters after upsamplers remove spectral images created by zero insertion. The design specifications for these filters directly impact the quality of the rate conversion and the overall system performance.

Decimation Process

The complete decimation process by factor M consists of lowpass filtering followed by downsampling. The anti-aliasing filter must have a cutoff frequency at or below π/M in normalized radian frequency, where π corresponds to half the input sample rate, to prevent aliasing when the rate is reduced. Any frequency content between π/M and π in the input signal will fold back into the baseband after downsampling unless removed by the filter.

The filter's passband should extend to the highest frequency of interest in the signal, while the stopband should attenuate potential aliasing components to acceptable levels. The transition bandwidth represents a compromise between filter complexity and spectral efficiency. Sharper transitions require higher-order filters with more coefficients and greater computational cost.

Stopband attenuation requirements follow from the noise floor the application must respect rather than from any fixed figure. The ideal signal-to-quantization-noise ratio of a full-scale sinusoid is about 6.02 dB per bit plus 1.76 dB, which places the floor near 98 dB for 16-bit audio and near 146 dB for 24-bit audio, so an audio decimation filter is specified to drive aliases beneath whichever floor applies. Communications systems instead derive stopband attenuation from adjacent-channel rejection requirements, and measurement instruments derive it from the accuracy claimed for the recorded data.

Efficient Decimation Implementation

Direct implementation of decimation wastes computation by evaluating filter outputs that will be discarded. For a decimation factor of M, M − 1 out of every M filter outputs need not be computed. Rearranging the computation to avoid this waste reduces the arithmetic rate, the multiplications per second, by the factor M. The work spent on each retained output is unchanged; what disappears is the work spent on the outputs that were going to be discarded.

The efficient approach computes only the retained outputs by stepping through the input signal M samples at a time. At each step, the filter's convolution sum is computed using only the input samples that contribute to that particular output. This restructuring requires tracking which input samples and filter coefficients participate in each output computation.

For an FIR filter with N coefficients operating at decimation factor M, the efficient implementation requires N multiplications and N − 1 additions per output sample. Since the output rate is M times lower than the input rate, this translates to N/M multiplications per input sample, compared to N multiplications per input sample for the naive approach. The savings become substantial for large decimation factors.

Multistage Decimation

Large decimation factors are most efficiently implemented as cascades of smaller factors. A 100:1 decimation, for example, might be realized as cascaded stages of 10:1, 5:1, and 2:1. The saving comes from the transition bands as much as from the rates. Each stage must attenuate only the bands that would fold onto the final passband, so every stage but the last can use a transition band far wider, and a filter far shorter, than the overall specification implies. The longest filter ends up in the last stage, where the sample rate is lowest and each coefficient therefore costs the least to evaluate.

The anti-aliasing filter at each stage need only prevent the aliasing produced by that stage. Its stopband may begin just below that stage's own output rate rather than just above the signal band, and only the final stage has to place its stopband edge close to the signal. The accumulated filtering across all stages provides the overall required frequency response.

A worked example shows the scale of the saving. Consider decimating a 10 kHz stream to 100 Hz, preserving content to 45 Hz, suppressing everything from 50 Hz upward, with a passband ripple of 1 percent and a stopband ripple of 0.1 percent. Kaiser's estimate for the number of coefficients N of an equiripple lowpass filter is

N ≈ (−10 log₁₀(δ_pδ_s) − 13) / (14.6 Δf / F) + 1

where δ_p and δ_s are the passband and stopband ripples, Δf is the transition width, and F is the sample rate at which the filter runs. A single stage must fit its 5 Hz transition into a 10 kHz input rate and needs about 5,070 coefficients; at one multiplication per coefficient for each of the 100 outputs per second, that is about 507,000 multiplications per second. Split as 10:1, then 5:1, then 2:1, the stages begin their stopbands 50 Hz below their own output rates, at 950 Hz, 150 Hz, and 50 Hz. Because the passband ripples of cascaded stages add in the worst case, each stage is given a third of the 1 percent budget, while each keeps the full 0.1 percent stopband ripple, since every frequency that could alias onto the signal lies in the stopband of at least one stage. The estimate then gives filters of about 33, 29, and 116 coefficients, costing 33,000, 5,800, and 11,600 multiplications per second, about 50,000 in all and a tenfold saving. Reversing the order to 2:1, then 5:1, then 10:1 raises the total to about 109,000 multiplications per second, because leaving the large factor until last forces the final filter, the one that must place its stopband edge just above the signal, to be designed at a high input rate, where the same transition width is a much smaller fraction of the band and demands five times as many coefficients. None of these counts exploits coefficient symmetry, which would roughly halve them all alike, and the estimate is a guide rather than a design: actual Parks-McClellan filters usually land within a few percent of it, which leaves the comparison unchanged.

The rule that falls out of such comparisons is to take the largest factor first in a decimation cascade. Beyond that, optimal factoring depends on the specific filter requirements, computational costs, and memory constraints. For power-of-two decimation factors, cascades of halfband filters are particularly efficient: every second coefficient of a halfband filter is zero apart from the center tap, and linear-phase symmetry halves the remainder again, so a halfband filter costs roughly a quarter of the multiplications of a conventional filter of the same length. Non-power-of-two factors require more general filter designs but can still benefit from cascade implementation.

CIC Decimators

Cascaded integrator-comb (CIC) filters provide extremely efficient decimation by large factors using only additions and subtractions, completely avoiding multiplication. The structure consists of N integrator stages operating at the input rate, a downsampler, and N comb stages operating at the output rate. The integrators accumulate running sums while the combs compute differences between delayed samples.

The structure behaves like N cascaded moving sums of RM input samples each, followed by the downsampler, where R is the decimation factor and M is the differential delay of the comb sections, counted in output samples and typically one or two. Each moving sum has the response of a Dirichlet kernel, which closely approximates a sinc function across the passband and the first few sidelobes, and the responses of the cascaded sums multiply. With the frequency f normalized to the input sample rate, the magnitude response of the N-stage decimator is therefore

|H(f)| = |sin(πRMf) / sin(πf)|^N

Within the input Nyquist band its nulls fall at multiples of the input rate divided by the product of the rate change and the differential delay, so a null sits at the center of every band that the decimation would fold onto the signal; with the usual unit differential delay, those centers, the multiples of the output rate, are the only nulls. The response peaks at zero frequency, where it equals (RM)^N. This response provides inherent anti-aliasing for the decimation operation, though with significant passband droop and limited stopband attenuation. Applications requiring better frequency response typically follow a CIC stage with compensation filtering.

CIC filters excel as the first decimation stage in high-rate systems where multiplications would be prohibitively expensive. A CIC stage might reduce the rate by a factor of 100 or more, after which conventional FIR filtering at the reduced rate provides the required frequency response. This combination achieves both efficiency and performance.

Register growth in CIC filters requires attention to prevent overflow. Each integrator accumulates values that can grow to large magnitudes before the comb stages produce the differences. Following Hogenauer's analysis, the worst-case gain through the decimator is its zero-frequency gain, (RM)^N, because the impulse response is never negative. A full-precision implementation must therefore size every register to the input width plus N log₂(RM) bits, rounded up, which for the common case of unit differential delay reduces to the input width plus N log₂(R). A four-stage decimator by 16 with unit differential delay and a 10-bit input therefore has a worst-case gain of 16⁴, or 65,536, and needs 26-bit registers. Provided every accumulator spans this range in two's-complement arithmetic, the integrators may overflow internally yet still yield correct results, because the comb differencing wraps modulo the same word length. Dropping bits from the top of any register breaks the output outright rather than degrading it gracefully; dropping them from the bottom only adds noise, as the discussion of fixed-point implementation below explains.

Interpolation Process

Upsampling by factor L inserts L − 1 zeros between each input sample, creating a signal at the higher rate with the same spectral content as the original but accompanied by L − 1 spectral images. These images are replicas of the baseband spectrum centered at multiples of the original sampling frequency. The interpolation filter removes these images, passing only the baseband component.

The interpolation filter's cutoff frequency should be π/L in the normalized frequency of the output rate, which places it at the original Nyquist frequency, suppressing the images while passing the signal. The filter's gain must be L, because a unity-gain filter would leave the interpolated signal at 1/L of the original amplitude: zero insertion spreads the original energy across L times as many samples, and the images the filter removes carry all but a fraction 1/L of it. This gain normalization ensures that the interpolated signal has the same amplitude as the original.

The quality of interpolation depends on how well the filter approximates the ideal lowpass response. Perfect reconstruction would require an ideal brick-wall filter with infinite duration, which is impractical. Real interpolation filters trade off complexity against image suppression and passband flatness, with the acceptable compromise depending on the application.

Efficient Interpolation Implementation

Direct implementation of interpolation multiplies each zero-inserted sample by filter coefficients, wasting computation on multiplications by zero. Since L − 1 out of every L input samples to the filter are zero, L − 1 out of every L terms in the filter's convolution sum contribute nothing to the output. Eliminating these unnecessary operations reduces computation by factor L.

The efficient implementation recognizes that only every Lth filter coefficient multiplies a non-zero input sample at any given output time. As the output index advances, a different one of L such subsets participates in the computation. The filter can be decomposed into L subfilters, each containing every Lth coefficient, with each subfilter producing one of the L output samples associated with each input sample.

This polyphase decomposition reduces the number of multiplications per output sample to N/L for an N-coefficient filter. Since the output rate is L times higher than the input rate, the total operations per input sample equals N, the same as filtering at the input rate without interpolation. The computational load grows with the filter order but not with the interpolation factor.

Interpolation Quality

Linear interpolation, the simplest approach, draws straight lines between adjacent samples. This corresponds to filtering with a triangular impulse response, which has poor frequency-domain characteristics including significant image content and passband droop. Linear interpolation is acceptable for visualization and some audio applications but inadequate for high-quality signal processing.

Sinc interpolation theoretically achieves perfect reconstruction by convolving with the sinc function, the impulse response of the ideal lowpass filter. However, the sinc function has infinite duration, requiring truncation for practical implementation. Windowed sinc interpolators achieve excellent quality with finite-length filters, with the window function controlling the trade-off between filter length and frequency response.

Polynomial interpolators use polynomials fitted to nearby samples to estimate intermediate values. Lagrange interpolators guarantee that the polynomial passes through the sample points. Splines trade exactness for smoothness in a subtler way: an interpolating cubic spline also passes through every sample and has a continuous second derivative, whereas the B-spline kernels often used directly as interpolation filters smooth the data instead of reproducing it unless the samples are prefiltered first. These methods can be implemented efficiently using Farrow structures that separate the time-invariant and time-varying parts of the computation.

The choice of interpolation method depends on quality requirements, computational constraints, and the specific application. Audio applications typically require high-quality interpolation to avoid audible artifacts. Communications systems may tolerate more image content if it falls outside the channel bandwidth. Real-time embedded systems may be limited to simple interpolators by processing power constraints.

Multistage Interpolation

Large interpolation factors benefit from multistage implementation, paralleling the efficiency gains of multistage decimation. Cascading smaller interpolation factors allows simpler filters at each stage, with the most complex filtering occurring at the lowest rate where computation is cheapest.

The ordering of stages matters, and it is the mirror of the decimation rule: the largest factor goes last. An interpolation cascade is a decimation cascade read backwards, stage for stage and cost for cost, so the 100:1 example above applies with the rates reversed. The first stage runs at the lowest rate and carries the long, sharp filter that fixes the final passband edge, while the last stage, running fastest, faces images already pushed far from the signal and needs only a short filter with a very wide transition band. Interpolating by 2, then 5, then 10 costs roughly half as much as the reversed order.

CIC interpolators provide multiplication-free rate increase for large factors, complementing CIC decimators. The structure reverses the decimator, with comb stages at the input rate followed by upsampling and integrator stages at the output rate. Compensation filtering at the lower rate addresses the passband droop and limited stopband rejection of the CIC response.

Polyphase Decomposition Concept

A filter with N coefficients can be decomposed into M polyphase components, each containing approximately N/M coefficients. The kth polyphase component contains coefficients h[k], h[k+M], h[k+2M], and so on, effectively subsampling the original impulse response. These components, when properly combined, reproduce the original filter's operation.

The decomposition exploits the structure of rate conversion operations. In decimation, only every Mth output is retained, meaning that each output depends on a specific pattern of input samples and coefficients. The polyphase view separates these patterns, computing only the needed outputs. In interpolation, zero-valued input samples contribute nothing to outputs, and the polyphase view eliminates the corresponding multiplications.

Mathematically, the z-transform of a filter H(z) can be expressed as the sum of M terms, each a polyphase component E_k(z) evaluated at z^M and delayed by z^−k:

H(z) = ∑_k=0^M−1 z^−k E_k(z^M)

This decomposition reveals the structure that enables efficient implementation and provides the algebraic framework for manipulating multirate systems.

Polyphase Decimation Structure

In polyphase decimation, the input signal is distributed to M branches, with each branch receiving every Mth sample starting from a different offset. Each branch filters its samples with one polyphase component of the anti-aliasing filter, producing a partial result at the reduced output rate. The partial results combine to form the decimated output.

This structure computes exactly the samples that will be retained, avoiding the wasted computation of naive implementations. Each polyphase filter operates at the output rate rather than the input rate, processing only the samples that affect the retained outputs. The total computation equals that of filtering at the output rate with a filter of the original length.

The input commutator that distributes samples to branches can be implemented as a circular buffer with multiple read pointers or as explicit demultiplexing logic. The choice depends on the implementation platform and the specific rate change factor. Hardware implementations often favor the commutator view, while software implementations may use pointer manipulation within a single buffer.

Polyphase Interpolation Structure

Polyphase interpolation reverses the decimation structure. Each polyphase component processes the input samples at the input rate, producing one of the L output samples associated with each input. An output commutator cycles through the branch outputs, assembling the full-rate interpolated signal from the partial contributions.

This structure eliminates multiplication by the zero-valued samples inserted during upsampling. Each polyphase filter contains only the coefficients that would multiply non-zero inputs at its corresponding output phase. The total computation is independent of the interpolation factor, depending only on the filter length and input rate.

The output commutator can be viewed as a time-division multiplexer selecting among the polyphase outputs. In hardware, this might be a literal multiplexer switching between branch outputs. In software, it might be a loop iterating through stored branch results or interleaved computation and output in a single loop.

Polyphase Filter Design

Designing polyphase structures starts with designing the prototype filter meeting the overall frequency response requirements. The prototype filter's order determines the complexity of each polyphase component, with longer filters providing better frequency response at the cost of more computation per component.

The prototype filter is then decomposed into polyphase components by distributing its coefficients across the branches. For decimation by M, the prototype has cutoff frequency π/M at the input rate, and the polyphase components are the M interleaved subsets of its coefficients. For interpolation by L, the prototype has cutoff frequency π/L at the output rate and is similarly decomposed into L branches.

Linear phase prototype filters yield polyphase components with favorable symmetry properties that can reduce computation further. If the prototype has symmetric coefficients, pairs of polyphase components are related by time reversal, and their inputs can be combined before filtering. This additional optimization approaches a factor of two reduction in multiplications.

Rational Rate Conversion

Converting between sample rates related by the rational factor L/M requires interpolation by L followed by decimation by M. The signal is first upsampled by L, inserting L − 1 zeros between each original sample. The result is then downsampled by M, retaining every Mth sample of the interpolated signal. The output rate equals (L/M) times the input rate.

The interpolation and decimation filters can be combined into a single filter operating between the input and output rates. This combined filter must satisfy both the interpolation requirement (removing images from upsampling) and the decimation requirement (preventing aliasing from downsampling). The cutoff frequency is the smaller of π/L and π/M, normalized to the intermediate rate of L times the input rate.

Polyphase implementation of rational rate conversion achieves efficiency by computing only the output samples actually needed. The polyphase structure cycles through L phases to produce the L intermediate samples per input sample, but only the phases corresponding to retained outputs need evaluation. For a combined filter of N coefficients at the intermediate rate, each output costs N/L multiplications, so the cost per input sample is N/M. For a given filter length the arithmetic rate therefore depends on M but not on L, although the filter itself must lengthen as L rises, since it runs at L times the input rate and must still achieve the same transition width in hertz.

The greatest common divisor of L and M should be factored out before implementation to minimize the number of polyphase branches. Converting 44100 Hz to 48000 Hz involves the ratio 160/147, not 48000/44100, because the greatest common divisor of 48000 and 44100 is 300. The reduced ratio calls for 160 polyphase branches rather than 48000. This reduction significantly impacts implementation complexity for rates with common factors.

Arbitrary Rate Conversion

When the ratio of input to output rate is irrational or unknown, arbitrary rate conversion techniques compute output samples at precisely specified time offsets from the input grid. Rather than predetermining which input samples and filter coefficients combine for each output, the system calculates these combinations dynamically based on the required output timing.

Polynomial-based interpolation computes output samples using weighted combinations of nearby input samples, with weights determined by the fractional time offset from the input grid. Lagrange interpolators, cubic interpolators, and higher-order polynomials provide increasingly accurate approximations of the ideal sinc interpolation at increasing computational cost.

The Farrow structure implements polynomial interpolation efficiently by separating time-invariant and time-varying computations. A bank of FIR filters computes polynomial basis functions of the input samples, and these results are combined using the fractional delay as a parameter. This structure enables efficient implementation of continuously variable fractional delay with fixed filter coefficients.

Asynchronous sample rate conversion uses arbitrary rate conversion to synchronize signals from unsynchronized sources. The conversion ratio may vary slowly as the source clocks drift relative to each other. A timing recovery system estimates the instantaneous rate ratio and adjusts the interpolator accordingly, maintaining synchronization despite clock imperfections.

Quality Considerations

Sample rate conversion quality is measured by how closely the converted signal matches ideal band-limited interpolation or decimation. Passband droop, image rejection, and aliasing rejection characterize the frequency-domain performance. Time-domain measures include step response overshoot, ringing, and group delay variation.

Audio sample rate conversion demands high quality to avoid audible artifacts. The governing criterion is that images and aliases fall beneath the noise floor of the material: the ideal quantization noise floor lies about 98 dB below full scale for 16-bit audio and about 146 dB below it for 24-bit audio, so a converter intended for professional work holds its stopband attenuation to a comparable depth and its passband flat within fractions of a decibel. Vendors of dedicated asynchronous sample rate converter parts quote dynamic range figures in the region of 140 dB, a vendor claim rather than an independent measurement. Consumer audio accepts somewhat lower specifications but still requires careful design to avoid audible degradation.

Communications systems balance quality against latency and complexity. Real-time systems may accept some degradation to meet timing constraints. Burst-mode systems converting between packet rates and continuous rates face additional challenges at packet boundaries that require special handling.

Testing sample rate converters requires appropriate metrics and test signals. Swept-sine testing reveals frequency response and image/alias levels. Multitone testing reveals intermodulation products that swept-sine testing might miss. Time-domain testing with impulses and steps reveals transient behavior and potential instabilities.

Analysis and Synthesis Structure

An M-channel analysis filter bank consists of M bandpass filters covering the frequency range of interest, each followed by downsampling by factor M. The analysis filters divide the spectrum into M bands, and the downsamplers reduce each band to the minimum rate necessary for its bandwidth. The resulting subband signals collectively represent the original signal at the same total sample rate.

The synthesis filter bank reverses this process. Each subband signal is upsampled by M, filtered to remove images, and summed with the other bands to reconstruct the full-band signal. The synthesis filters must complement the analysis filters to achieve accurate reconstruction of signals that pass through the system unprocessed.

The analysis-synthesis cascade can process each subband independently before reconstruction. Different processing can be applied to different frequency bands, enabling operations like frequency-dependent gain adjustment, selective noise reduction, or subband coding. The filter bank structure isolates these operations from each other, simplifying design and implementation.

Perfect Reconstruction

A filter bank achieves perfect reconstruction if the analysis-synthesis cascade reproduces the input signal exactly, possibly with a fixed delay. This ideal behavior requires precise relationships between the analysis and synthesis filters that cancel all distortions introduced by the subband processing, including aliasing from the downsamplers.

Perfect reconstruction conditions can be expressed in terms of the filter transfer functions. The overall transfer function must be a pure delay (or a scaled delay for systems with gain), and the aliasing terms introduced by downsampling must cancel exactly. These conditions impose constraints that link the analysis and synthesis filters together.

For critically sampled filter banks, where the decimation factor equals the number of channels, perfect reconstruction requires specific filter designs. The constraints are stringent, limiting the achievable frequency selectivity for a given filter order. Oversampled filter banks relax these constraints by using smaller decimation factors, trading rate efficiency for design flexibility.

Near-perfect reconstruction accepts small reconstruction errors in exchange for simpler filter designs or better frequency selectivity. Audio coding applications often use near-perfect reconstruction filter banks where the errors are below audible thresholds. The acceptable error level depends on the application's perceptual or measurement requirements.

Polyphase Filter Bank Implementation

Polyphase structures provide efficient filter bank implementation by exploiting the common elements across channels. The analysis filter bank can be restructured as a polyphase network followed by a transform operation, typically the discrete Fourier transform (DFT). This structure reduces computation by sharing the polyphase filtering across all channels.

The DFT modulated filter bank uses a single prototype lowpass filter, modulated to different frequencies to create the analysis filters. The polyphase components of the prototype filter appear in the polyphase network, and the DFT provides the modulation. The fast Fourier transform (FFT) computes the DFT efficiently, making this structure attractive for large numbers of channels.

Synthesis filter banks have dual polyphase structures, with the inverse DFT followed by polyphase filtering. The synthesis prototype filter may differ from the analysis prototype, depending on the perfect reconstruction requirements. The overall structure achieves efficient implementation of both analysis and synthesis with shared computational resources.

Two-Channel QMF Structure

The two-channel QMF bank splits a signal into low-frequency and high-frequency bands using complementary filters. The lowpass filter H₀(z) passes frequencies below π/2, while the highpass filter H₁(z) passes frequencies above π/2. Both bands are downsampled by 2, and the synthesis bank upsamples and filters to reconstruct the original signal.

The quadrature mirror relationship requires H₁(z) = H₀(−z), which in the time domain alternates the sign of every other lowpass coefficient and so shifts the response by π. This relationship ensures that the two filters together cover the full spectrum without gaps. The frequency responses are mirror images about π/2, hence the quadrature mirror name.

Aliasing cancellation in QMF banks requires specific relationships between analysis and synthesis filters. In the classical two-channel structure the synthesis filters are the analysis filters themselves, with the highpass negated: F₀(z) = H₀(z) and F₁(z) = −H₁(z). Time reversal belongs to the conjugate quadrature designs described below, not here. With this choice the two aliasing terms are equal and opposite and cancel exactly, leaving an overall transfer function of ½[H₀²(z) − H₀²(−z)].

Once aliasing has been cancelled, perfect reconstruction demands that this remaining transfer function be a pure delay, and for a linear-phase FIR lowpass that is possible only in the trivial two-tap case. The Haar pair is the one exception: with both coefficients equal to 1/√2, the transfer function collapses to a single delay, giving exact reconstruction with exact linear phase, though with frequency selectivity far too poor for subband coding. Any longer symmetric prototype leaves residual amplitude distortion, which is why practical designs choose between nearly perfect reconstruction with linear phase, as in Johnston's classic filters, and exact reconstruction with nonlinear phase.

Conjugate Quadrature Filters

Conjugate quadrature filters (CQF), also called power-symmetric or orthogonal filters, achieve perfect reconstruction in two-channel filter banks. The CQF relationship requires that the sum of squared magnitude responses of lowpass and highpass filters equals a constant. This power complementarity ensures that signal energy is preserved through the analysis-synthesis process.

CQF banks derive the highpass from the lowpass by modulation and time reversal. In the time domain the relationship is h₁[n] = (−1)ⁿh₀[N − n], where N is the filter order, which in the z domain reads

H₁(z) = −z^−NH₀(−z⁻¹)

and the synthesis filters equal the time-reversed analysis filters. The condition that the relationship alone does not show is decisive: the two aliasing terms cancel only when N is odd, that is, when the filters have an even number of taps. That is no accident of convention. Power symmetry forces the lowpass autocorrelation to vanish at every even lag except zero, and for an odd-length filter the autocorrelation at lag N is the product of the two end taps, which cannot vanish unless the filter is really shorter. An overall sign on the highpass, by contrast, is a matter of convention, because it appears in the analysis and synthesis filters alike and cancels. This configuration achieves exact reconstruction with FIR filters but requires nonlinear phase for the individual filters.

Orthogonal wavelets, including the Daubechies family, arise from iterated CQF filter banks. The perfect reconstruction property of the filter bank guarantees that the wavelet transform is invertible. The orthogonality property preserves energy, so the sum of the squared transform coefficients equals the sum of the squared samples and the inverse transform is the transpose of the forward one. Orthogonal wavelet bases also tend to decorrelate signals whose energy concentrates at low frequencies, which is what makes them useful for compression, although how far they do so depends on the signal statistics rather than on orthogonality alone.

Biorthogonal Filter Banks

Biorthogonal filter banks use different filters for analysis and synthesis, relaxing the orthogonality constraint to enable linear phase and symmetric filters. The biorthogonality condition requires that the analysis filters be orthogonal to shifted versions of the synthesis filters, but the analysis and synthesis sets need not be orthogonal within themselves.

Linear phase biorthogonal filters are achievable with symmetric or antisymmetric coefficients. These filters produce symmetric wavelets that avoid the phase distortion of orthogonal designs. The 5/3 and 9/7 wavelet filters used in JPEG 2000 are biorthogonal designs with linear phase and strong coding performance. The 5/3 pair, implemented as an integer-to-integer lifting transform, serves the reversible, lossless path, while the 9/7 pair serves the irreversible, lossy path.

Biorthogonal filter design offers more degrees of freedom than orthogonal design, enabling optimization for specific applications. Regularity, which controls the smoothness of the associated wavelet, can be maximized independently for analysis and synthesis. Frequency response can be optimized separately for the lowpass and highpass filters within constraints required for perfect reconstruction.

Conditions for Perfect Reconstruction

Perfect reconstruction in an M-channel filter bank requires that the cascade of analysis and synthesis operations produces a pure delay or scaled delay. Mathematically, if X(z) is the input and Y(z) is the output, perfect reconstruction means Y(z) = c z^−dX(z) for some constant c and integer delay d. All other terms, including aliasing, must cancel exactly.

The aliasing cancellation conditions constrain the relationship between analysis and synthesis filters. In a critically sampled M-channel bank, M − 1 aliasing terms arise from the downsampling operation. Each must be eliminated by appropriate choice of synthesis filters given the analysis filters, or by joint design of both sets.

The distortion function, representing the gain and phase through the system in the absence of aliasing, must be a pure delay for perfect reconstruction. This function depends on both analysis and synthesis filters and equals a delay only for specific filter combinations. Deviation from a pure delay indicates amplitude distortion or phase distortion in the reconstructed signal.

Paraunitary Filter Banks

Paraunitary filter banks represent the analysis filter bank as a paraunitary matrix, meaning that the matrix times its paraconjugate, the transposed matrix with each coefficient conjugated and z replaced by its reciprocal, equals a constant times the identity matrix. This algebraic structure guarantees perfect reconstruction and ensures that the filter bank preserves signal energy. Orthogonal filter banks are paraunitary.

The paraunitary condition provides a framework for designing and analyzing perfect reconstruction systems. Filter banks can be constructed from cascades of simple paraunitary building blocks, with each block guaranteed to preserve the paraunitary property. This modularity simplifies both design and implementation.

Lattice structures implement paraunitary filter banks using cascades of rotation matrices, each parameterized by a single angle. The angles provide independent design parameters that can be optimized for frequency response or other criteria. The resulting filters automatically satisfy the perfect reconstruction conditions regardless of the angle values.

Design Methods for Perfect Reconstruction

Direct design methods specify the desired filter frequency responses and solve for coefficients satisfying the perfect reconstruction constraints. This approach works for short filters but becomes computationally challenging as filter length increases. The nonlinear constraints make the optimization problem difficult, and multiple solutions may exist with different frequency response properties.

Lattice-based design parameterizes the filter bank by rotation angles and optimizes these parameters for the desired response. Since any angle values yield perfect reconstruction, the optimization focuses purely on frequency response without constraints. This approach scales well to long filters and automatically satisfies reconstruction requirements.

Lifting-based design constructs filter banks through sequences of prediction and update steps. Starting from a simple filter bank like the Haar, lifting steps incrementally modify the frequency response while preserving perfect reconstruction. This approach provides intuition about the design process and enables integer-to-integer transforms useful for lossless coding.

Digital Audio Processing

Audio sample rate conversion enables interoperability between equipment operating at different standard rates. Professional audio uses 48 kHz and 96 kHz, while consumer formats use 44.1 kHz (CD) and various compressed formats. High-quality sample rate converters maintain audio fidelity when transferring between systems or combining sources at different rates.

Oversampling in digital-to-analog conversion uses interpolation to push the spectral images far above the audio band, where simple analog filters can remove them. Rather than requiring sharp analog filters at the audio Nyquist frequency, oversampling allows gentle analog filtering that introduces less phase distortion. Combined with noise shaping, which moves the converter's own quantization noise out of the audio band, the technique enables high-quality audio from relatively simple analog output stages.

Filter banks underlie the perceptual audio compression standards. MP3 uses a hybrid structure in which a 32-band polyphase quadrature filter bank feeds a modified discrete cosine transform, while AAC applies the modified discrete cosine transform directly. Either way the audio arrives as frequency bands that can be quantized according to psychoacoustic models of human hearing. Bands where masking reduces audibility receive coarser quantization, achieving dramatic compression while maintaining perceptual quality.

Communications Systems

Digital down-conversion in software-defined radios combines frequency translation with decimation to bring a signal from a high intermediate frequency to baseband at a sample rate matched to its bandwidth. A complex mixer shifts the desired signal to zero frequency, and cascaded decimation stages reduce the sample rate while filtering out adjacent channels. This digital approach replaces analog mixing and filtering with flexible software implementations.

Channelization in multi-channel receivers uses filter banks to separate signals occupying adjacent frequency bands. A DFT-modulated filter bank efficiently extracts many channels from a wideband digitized input, enabling simultaneous reception of multiple signals with a single analog front end. This approach finds application in cellular base stations, spectrum monitoring, and satellite communications.

Symbol timing recovery in digital receivers uses interpolation to sample the received signal at optimal points within each symbol period. The timing recovery loop estimates the optimal sampling instant and adjusts the interpolator to sample there, even when the transmitter and receiver clocks are not synchronized. A polyphase bank of fractional-delay filters supplies a fixed set of finely spaced delays, and a Farrow structure supplies a genuinely continuous one where the residual timing error must be driven to zero.

Image and Video Processing

Image scaling requires two-dimensional sample rate conversion to change image resolution. Separable processing applies one-dimensional interpolation or decimation first horizontally, then vertically. The filter design must consider the specific artifacts visible in images, including aliasing that appears as jagged edges and ringing around sharp transitions.

Video format conversion between different frame rates and resolutions uses sophisticated multirate techniques, though the frame-rate case is not a resampling problem in the ordinary sense. Converting film at 24 frames per second to interlaced video near 30 frames per second uses 3:2 pulldown, which spreads four film frames over five video frames by drawing three video fields from one film frame and two from the next. No frame is discarded and no new frame is computed; fields are repeated, at the cost of the uneven motion cadence known as judder. For broadcast at 29.97 frames per second the film is additionally run at 23.976 frames per second, slower by one part in a thousand. Resolution conversion for display on different devices does involve genuine resampling to match the display pixel count, and motion-compensated rate conversion improves quality by synthesizing new frames along estimated motion trajectories rather than repeating existing ones.

Wavelet-based image compression in JPEG 2000 uses two-channel filter banks applied iteratively to decompose images into multiple resolution levels. The decomposition concentrates image energy in a small number of coefficients, enabling efficient compression. The multiresolution structure supports progressive transmission where image quality improves as more data arrives.

Instrumentation and Measurement

Digital oscilloscopes use decimation to match the stored sample rate to the display resolution and measurement needs. High-speed acquisition captures signals at rates sufficient to resolve the fastest features, while decimation reduces the data to manageable amounts for display and analysis. Variable decimation enables zoom functions that show different time scales.

Spectrum analyzers use filter banks to compute frequency-domain representations of signals. An FFT, itself a uniform filter bank, trades frequency resolution against time resolution, because a finer resolution needs a longer record. Other filter bank designs can provide better control over this trade-off, with potential for non-uniform frequency resolution matched to the analysis requirements.

Biomedical signal processing often operates on signals of very low bandwidth. Routine clinical electroencephalography records content below roughly 100 Hz, and IEC 60601-2-25, the particular standard for diagnostic electrocardiographs, requires a frequency response equivalent to a passband of 0.05 Hz to 150 Hz, which the American Heart Association recommends extending to 250 Hz for children. Decimation reduces high sample rate inputs to rates appropriate for these bandwidths, reducing storage and processing requirements. Multirate processing enables real-time analysis on resource-constrained portable devices.

Fixed-Point Implementation

Fixed-point arithmetic requires careful management of word lengths throughout the multirate system. Filter coefficients must be quantized to the available precision, affecting frequency response accuracy. Signal samples experience quantization at each processing stage, accumulating quantization noise that degrades signal-to-noise ratio.

Rate changes move the point at which dynamic range is at risk rather than changing the sample values themselves. Decimation leaves in-band signal levels unchanged, because the anti-aliasing filter has unity passband gain and the downsampler only discards samples, although filtering can still carry individual outputs past the input's full scale: removing the harmonics of a full-scale square wave leaves a fundamental 4/π times as large, and a sharp filter overshoots on transients. What needs the most headroom is the accumulator that forms the convolution sum, which must hold a sum of many products. Interpolation is the case that genuinely alters levels, since zero insertion and image removal together leave the signal at 1/L of its original amplitude and the interpolation filter carries a compensating gain of L. Where that gain is applied decides the outcome: too late and the intermediate samples waste dynamic range, too early and they overflow.

CIC filters require special attention to register sizing because of their recursive structure. The integrators accumulate values that grow without bound under a steady input, so the most significant bit of every register must sit where the worst-case gain described earlier places it, for the two's-complement wraparound to resolve correctly at the comb outputs. The word length need not be uniform, however: Hogenauer's analysis also shows how many least significant bits each stage may discard, and the later stages may discard more, because truncation noise injected late in the chain passes through less of the filter's remaining gain than noise injected early. Proper sizing throughout the structure ensures correct operation with minimum hardware cost.

Memory Organization

Circular buffers efficiently manage the delay lines required for filtering and rate conversion. Rather than shifting data at each sample, pointer arithmetic tracks the current position in a fixed memory region. This approach eliminates data movement overhead and is particularly important for long filters or high sample rates.

Polyphase structures partition memory according to the processing phases. Input samples can be distributed to separate buffers for each polyphase branch, or a single buffer can be accessed with phase-dependent indexing. The choice depends on the memory architecture and whether parallel or sequential processing is used.

Buffer management between processing stages with different rates requires careful design. The rate mismatch means that buffer fill and drain rates differ, requiring sizing and flow control to prevent overflow or underflow. Double-buffering or more complex schemes may be needed to decouple stages with different timing requirements.

Real-Time Operation

Real-time multirate systems must complete all processing within the time available at the highest sample rate. Although multirate techniques reduce average computational load, worst-case timing must still meet deadlines. Cascaded rate changes create variable computational loads that complicate scheduling analysis.

Latency in multirate systems includes contributions from filter delays at each rate and buffering between stages. The group delay of anti-aliasing and interpolation filters translates to signal delay in the output. Applications sensitive to latency must account for these delays, potentially accepting reduced filter quality to minimize delay.

Interrupt-driven processing matches well to multirate systems, with different processing triggered at different rates. High-rate processing handles simple operations like CIC filtering, while lower-rate interrupts perform more complex operations with more time available per sample. Giving the higher-rate tasks the higher priorities, as rate-monotonic scheduling does, protects their tighter deadlines.