Sampling, Quantization, and Signal Reconstruction

Introduction

Digital signal processing begins by converting a physical or continuous-time signal into a sequence of numbers that can be stored, transmitted, and manipulated by digital hardware. This conversion is not a single operation. It consists principally of sampling, which discretizes time; quantization, which discretizes amplitude; and encoding, which represents quantized values using binary words. Reconstruction performs the reverse operation: it produces a continuous-time waveform from discrete samples.

These operations determine the information preserved by a measurement system and the distortions introduced by it. In communications, instrumentation, control, radar, medical imaging, audio, and machine learning hardware, inappropriate sampling or quantization can cause irreversible errors. At postgraduate level, it is therefore important to understand not only the Nyquist criterion, but also its assumptions, spectral interpretation, reconstruction filters, quantization noise, oversampling, aliasing, and practical nonidealities.

Learning Objectives

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

1. distinguish continuous-time, discrete-time, discrete-amplitude, and digital signals;
2. derive the spectral replication caused by ideal impulse-train sampling;
3. state and apply the sampling theorem, including its assumptions and limitations;
4. explain aliasing and select appropriate anti-aliasing filters;
5. analyze uniform quantization, quantization error, signal-to-quantization-noise ratio, and dynamic range;
6. describe reconstruction using ideal and practical digital-to-analog conversion;
7. evaluate oversampling, undersampling, aperture effects, jitter, and finite-word-length effects;
8. design a sampling and quantization chain for an engineering application.

1. Signal Representation and the Conversion Chain

A continuous-time signal x(t) is defined for every real time t and may have continuous amplitude. A sampled signal consists of values x[n] = x(nT_s), where n is an integer and T_s is the sampling interval in seconds. The sampling frequency is

where f_s is measured in hertz. A sampled signal is discrete in time but may still have arbitrary amplitude. Quantization maps each sample to one of a finite set of amplitude levels. The resulting signal is discrete in both time and amplitude; binary encoding then produces a digital word.

A typical acquisition chain is:

sensor → analogue conditioning → anti-alias filter → sampler/sample-and-hold → quantizer/ADC → digital processing

A reconstruction chain is:

digital data → DAC → zero-order hold → reconstruction filter → analogue output

Figure: Principal stages of analogue-to-digital conversion and digital-to-analogue reconstruction.

2. Ideal Sampling and the Sampling Theorem

Ideal uniform sampling can be represented by multiplying x(t) by an impulse train

where δ(t) is the Dirac impulse. The sampled signal is

The Fourier transform of the impulse train is also an impulse train. Using frequency f in hertz, the spectrum of the sampled signal is

Thus sampling creates repeated copies of the original spectrum, separated by f_s and scaled by 1/T_s. This relation assumes ideal impulses, exact uniform timing, and a signal whose Fourier transform exists in the relevant sense.

The band-limited sampling theorem

If x(t) is strictly band-limited to |f| ≤ B Hz, and the spectral replicas do not overlap, perfect recovery is possible when

The quantity 2B is the Nyquist rate. The frequency f_s/2 is the Nyquist frequency of a sampler operating at f_s. The theorem is a sufficient condition for exact reconstruction, not a universal prescription for every practical system. Equality is unsafe for spectra containing energy exactly at the boundary, and real filters require a transition band; consequently, practical designs use margin above 2B.

The condition can also be expressed for angular frequency. If the highest angular frequency is Ω_B radians per second, then ω_s > 2Ω_B, where ω_s = 2π/T_s. Mixing hertz and radians per second without the factor 2π is a common error.

Reconstruction by sinc interpolation

For an ideal band-limited signal sampled above the Nyquist rate, reconstruction is

where sinc(u) = sin(πu)/(πu). Each sample is weighted by a shifted sinc function. This ideal interpolation extends infinitely in time and is therefore not exactly realizable. Practical systems approximate it with finite-order filters and oversampling.

3. Aliasing

Aliasing occurs when spectral components above the Nyquist frequency appear after sampling as lower-frequency components. For a sinusoid x(t) = A cos(2πf_0t + φ), sampled at f_s, frequencies that differ by an integer multiple of f_s produce the same sample sequence:

where k is an integer chosen so that f_alias lies in [0, f_s/2]. The amplitude A and phase representation may also be affected according to the chosen equivalent frequency.

For example, a 9 kHz sinusoid sampled at 12 kHz appears at |9 − 12| = 3 kHz. No digital operation performed after sampling can determine that the original was 9 kHz rather than 3 kHz, because the samples are identical under ideal conditions.

Aliasing is irreversible. The correct remedy is analogue filtering before the sampler. An anti-aliasing filter attenuates energy above the permitted input band. Its design must account for passband edge, stopband edge, allowable ripple, required attenuation, and the available sampling frequency. A guard band between the signal band and f_s/2 reduces filter complexity.

Band-pass sampling

A low-pass signal is not the only signal that can be sampled. A band-pass signal occupying [f_L, f_H] may be sampled below 2f_H if its spectral replicas do not overlap. This is called band-pass, harmonic, or undersampling. The selected sampling frequency must satisfy replica-separation inequalities and must preserve the intended spectral order. It is not safe to apply the simple low-pass rule mechanically; the entire band and filter transition must be considered.

4. Practical Sampling Nonidealities

Sample-and-hold

A practical ADC usually holds the input approximately constant during conversion. A zero-order hold has impulse response extending over one sampling interval and frequency response

with the normalized sinc definition above. The magnitude has a sin(x)/x droop and a phase delay. Equalization may compensate moderate droop, but it also amplifies noise near the band edge if used excessively.

Aperture time and aperture uncertainty

The aperture is the interval during which the input is observed. Finite aperture averages a rapidly varying signal and causes amplitude loss. Timing uncertainty, or jitter, produces voltage error approximately

for a sinusoidal input, when jitter is small and σ_t is the rms timing uncertainty. Here f_in is input frequency in hertz, V_rms is signal rms voltage, and σ_t is rms timing error in seconds. The corresponding jitter-limited signal-to-noise ratio is approximately

This shows why clock quality becomes increasingly important at high input frequencies.

Nonuniform sampling

If sampling instants are not exactly nT_s, the simple spectral-replica model must be modified. Timing errors can be modeled as t_n = nT_s + ε_n. Their effect depends on whether ε_n is deterministic, periodic, or random. Calibration, low-jitter clock distribution, and clock-recovery design are therefore central in high-speed converters.

5. Quantization

Quantization replaces each sample by a member of a finite amplitude set. For a uniform quantizer with input range from −V_F to +V_F and N bits, the number of levels is

and the quantization step is

assuming the stated range is the full-scale peak-to-peak range. A mid-rise or mid-tread quantizer assigns samples to intervals of width ∆. The quantization error is

For a rounding quantizer that is not overloaded, −∆/2 ≤ e_q[n] ≤ ∆/2. This bound is deterministic. Treating quantization error as white noise requires additional assumptions: the signal should sufficiently exercise many quantizer levels, the error should be weakly correlated with the signal, and overload should be absent. These assumptions fail for small periodic or slowly varying inputs.

If the error is modeled as uniformly distributed over [−∆/2, ∆/2], its variance is

and its rms value is ∆/√12. For a full-scale sinusoid with peak amplitude V_F, the ideal quantization signal-to-noise ratio is

This result assumes an ideal N-bit converter, a full-scale sinusoid, rounding quantization, no overload, and quantization noise spread over the Nyquist band. For a signal using only a fraction of full scale, the SNR is lower.

Quantization overload and effective resolution

If an input exceeds the converter range, clipping occurs. The error is then not bounded by ±∆/2 and may generate strong harmonics. Headroom is needed for signal peaks and transients. Real converters are characterized by effective number of bits (ENOB), which includes quantization noise, thermal noise, distortion, clock jitter, reference noise, and circuit imperfections. A common estimate from measured SINAD is

where SINAD is in decibels and the sinusoidal full-scale assumptions approximately apply.

Dither

Dither is deliberately added noise used to decorrelate quantization error from the input. It can eliminate deterministic distortion at the expense of increased noise. Properly designed dither is especially useful in audio, precision measurement, and low-level signal processing. It does not increase the physical resolution of the converter; it changes the character and statistical behavior of the error.

6. Oversampling, Noise Shaping, and Multirate Conversion

Oversampling uses a sampling rate substantially higher than the minimum required. It relaxes the analogue anti-alias filter, provides transition-band margin, and can improve in-band noise performance when quantization noise is distributed over a wider Nyquist band. If white quantization noise is spread over the Nyquist bandwidth and a digital filter retains only a fraction of that bandwidth, the in-band noise is reduced.

Noise-shaping converters, including delta-sigma architectures, use feedback to move quantization noise toward high frequencies, where digital filtering removes it. Their advantages include high resolution over a selected band and relaxed analogue filtering; their trade-offs include latency, loop-stability constraints, out-of-band noise, and dependence on a decimation filter.

Decimation reduces the sample rate after low-pass filtering. Interpolation increases it by inserting samples, usually zeros, followed by low-pass filtering. Filtering before rate reduction is essential: otherwise spectral components above the new Nyquist frequency alias into the retained band. In a rational rate converter with output/input rate ratio M/L, filtering and polyphase implementation can greatly reduce computation.

7. Signal Reconstruction and Digital-to-Analogue Conversion

Reconstruction is the estimation of a continuous-time waveform from samples. The ideal operation is low-pass interpolation with cutoff between the signal bandwidth and the first spectral image. The ideal brick-wall filter is noncausal and infinitely long, so practical systems use finite impulse response (FIR), infinite impulse response (IIR), or analogue reconstruction filters.

A DAC commonly implements a zero-order hold, producing a staircase waveform. The reconstruction filter removes staircase images around integer multiples of f_s. Its design should specify:

• passband edge and allowable amplitude error;
• stopband beginning and required image attenuation;
• phase or group-delay requirements;
• output settling time and latency;
• tolerance, temperature, and load effects.

The reconstructed output is not necessarily the unique original signal unless the band-limiting assumptions hold. If the original contained out-of-band energy that was not filtered before sampling, reconstruction cannot recover it or separate it from aliases.

8. Worked Design Example: Measurement of a Vibration Signal

Consider an accelerometer system whose useful vibration band is 0–18 kHz. The required passband amplitude error is at most 0.2 dB, and frequencies above 24 kHz must be attenuated by at least 80 dB before conversion. The ADC has a ±2 V input range and 14-bit resolution.

A sampling rate of exactly 36 kS/s is theoretically sufficient for a strictly band-limited signal, but it leaves no transition band between 18 kHz and the Nyquist frequency. Select f_s = 64 kS/s, giving a Nyquist frequency of 32 kHz. The analogue anti-alias filter must be approximately flat to 18 kHz and provide 80 dB attenuation by 24 kHz. The transition width is 6 kHz, making the filter substantially more realizable than one designed at 36 kS/s.

The ADC has L = 2¹⁴ = 16,384 levels. Its full-scale peak-to-peak range is 4 V, so

For an ideal full-scale sinusoid, the theoretical quantization SNR is

This is an upper bound. Actual SINAD will be lower because of sensor noise, analogue filter noise, reference instability, distortion, and clock jitter. If the vibration signal normally occupies only one quarter of full-scale peak amplitude, its ideal SNR is approximately 12 dB lower than the full-scale value, because the rms signal amplitude is reduced by a factor of four in power ratio.

After digital processing, if the required output bandwidth is only 10 kHz, a digital low-pass filter can be followed by decimation, for example from 64 kS/s to 32 kS/s, provided filtering sufficiently attenuates content above 16 kHz before the rate reduction. The final reconstructed signal should pass through a DAC and analogue smoothing filter if an analogue actuator or monitoring instrument is driven.

The design illustrates a central principle: sampling rate, analogue filter order, quantizer range, resolution, clock quality, and downstream processing must be selected as one system rather than independently.

9. Illustrative Case: An Industrial Condition-Monitoring System

Background. A factory monitors a motor using accelerometers. Engineers initially sample at 20 kS/s because the dominant vibration component appears near 7 kHz.

Problem. A newly installed gearbox produces components near 13 kHz and occasional resonances near 16 kHz. The 20 kS/s sampler has a 10 kHz Nyquist frequency. The 13 kHz component aliases to 7 kHz, overlapping the original motor vibration and causing false alarms.

Analysis. The digital spectrum alone cannot distinguish the true 7 kHz component from the aliased 13 kHz component. Investigation shows that the analogue input filter has insufficient attenuation above 10 kHz. The failure is therefore not a software classification problem; information was lost before digital processing.

Approach. The team changes the acquisition rate to 51.2 kS/s, designs an anti-alias filter with a passband through 18 kHz and a stopband beginning at 23 kHz, and validates attenuation using swept-frequency tests. A digital decimation stage then produces 12.8 kS/s data for the diagnostic algorithm after low-pass filtering. Clock jitter, sensor saturation, and ADC overload are tested under maximum motor load.

Outcome. The 13 kHz component remains distinguishable, false alarms are reduced, and the lower-rate diagnostic stream is retained for storage. The higher-rate stream is preserved selectively for event investigation.

Learning points. Aliasing cannot be removed after conversion; anti-alias filtering must precede sampling; a convenient nominal sample rate may be inadequate when operating conditions change; and validation must include analogue hardware, clock behavior, converter range, and digital filtering.

10. Design Trade-offs, Governance, and Contemporary Practice

Higher sampling rates increase data volume, memory, communication bandwidth, and computational load. More bits improve amplitude resolution but may increase power, conversion time, and cost. Wider analogue bandwidth may capture useful information but also admits interference. Engineering specifications should therefore state the required information bandwidth, amplitude accuracy, dynamic range, latency, allowable distortion, and uncertainty—not merely the ADC bit count.

In safety-critical and medical systems, traceability of calibration, timing, saturation detection, and data integrity is essential. Raw samples should not be silently discarded when they are needed for diagnosis or forensic analysis. Metadata should record sampling rate, filter settings, gain, calibration date, units, and overload indicators. In scientific and industrial AI systems, data preprocessing must not introduce undocumented aliasing or resampling artifacts that bias model training. Reproducible pipelines should preserve the conversion configuration and distinguish measured data from reconstructed or interpolated data.

Best practice includes:

1. define the highest information frequency before choosing f_s;
2. provide an analogue anti-alias filter with a documented transition band;
3. leave practical margin above the theoretical Nyquist rate;
4. prevent overload and use appropriate headroom;
5. evaluate ENOB, SINAD, jitter, aperture, and filter tolerances;
6. low-pass filter before decimation and image-filter after interpolation;
7. verify both time-domain waveforms and spectra using calibrated test signals;
8. document units, scaling, latency, and all resampling operations.

Challenges and Limitations

The ideal sampling theorem assumes perfect band limitation, exact timing, ideal impulses, unlimited computation, and perfect filters. Physical signals are rarely strictly band-limited, and real filters have finite transition bands. Quantization noise is not always white or independent of the signal. Converter nonlinearity, missing codes, thermal noise, reference noise, crosstalk, and electromagnetic interference may dominate the ideal quantization limit. Reconstruction is an estimation problem whenever the original signal violates the band-limit assumption. These limitations should be treated as design parameters rather than ignored idealizations.

Chapter Summary

Sampling discretizes time, while quantization discretizes amplitude. Ideal uniform sampling replicates the signal spectrum at integer multiples of the sampling frequency. A strictly band-limited signal with highest frequency B can be recovered when f_s > 2B, but practical systems require transition-band margin and analogue anti-alias filtering. Aliasing is irreversible after sampling.

Uniform N-bit quantization provides 2^N levels and, under ideal full-scale sinusoidal assumptions, an approximate SNR of 6.02N + 1.76 dB. The quantization-error model is useful only when its statistical assumptions are reasonable; overload, correlation, and converter nonidealities can invalidate it. Dither, oversampling, and noise shaping modify the distribution or spectral placement of quantization noise.

Reconstruction uses interpolation and low-pass filtering. DAC zero-order holds create images and sin(x)/x droop, which practical reconstruction filters must address. Sound engineering design integrates bandwidth, sampling rate, filter transition, quantizer range, resolution, clock jitter, data rate, latency, and governance requirements. The central professional lesson is that acquisition and reconstruction form a complete system: no later digital algorithm can recover information destroyed by inadequate sampling or quantization.

Key Terms

Sampling: Converting a continuous-time signal into values at selected time instants.

Sampling interval: Time between consecutive uniform samples, T_s.

Sampling frequency: Number of samples per second, f_s = 1/T_s.

Nyquist rate: Twice the highest frequency of a strictly band-limited low-pass signal.

Nyquist frequency: Half the sampling frequency.

Aliasing: Irreversible spectral overlap caused by insufficient sampling.

Anti-aliasing filter: Analogue filter placed before an ADC to attenuate unwanted high frequencies.

Quantization: Mapping continuous-valued samples to a finite set of amplitudes.

Quantization step: Amplitude spacing between adjacent uniform quantizer levels.

Quantization error: Difference between the input sample and its quantized value.

Overload: Operation beyond converter input range, usually causing clipping.

Dither: Intentional noise added to reduce signal-correlated quantization distortion.

Oversampling: Sampling substantially faster than the minimum theoretical rate.

Noise shaping: Feedback-based spectral redistribution of quantization noise.

Decimation: Filtering and reducing a sampling rate.

Interpolation: Increasing a sampling rate, normally followed by low-pass filtering.

Zero-order hold: DAC output stage that holds each sample value for one sampling interval.

Reconstruction filter: Filter that removes spectral images and smooths a DAC output.

Aperture jitter: Uncertainty in the actual sampling instant.

ENOB: Effective number of bits inferred from measured converter performance.

Review Questions

Short Answer Questions

1. Distinguish a sampled signal from a quantized signal.
2. State the low-pass sampling theorem and its principal assumption.
3. Why must anti-alias filtering occur before the ADC?
4. Define quantization overload.
5. What function does a reconstruction filter perform after a DAC?

Long Answer Questions

1. Derive the spectrum of an impulse-train-sampled signal and explain the role of spectral replicas.
2. Explain uniform quantization, derive the variance ∆²/12 under the uniform-error model, and state when that model is inappropriate.
3. Compare ideal sinc reconstruction with practical zero-order-hold reconstruction.
4. Discuss oversampling, decimation, interpolation, and noise shaping as a multirate signal-processing framework.
5. Analyze the effects of aperture jitter, converter range, ENOB, and analogue filter design on measurement accuracy.

Analytical / Application Questions

1. A 14 kHz sinusoid is sampled at 20 kS/s. Determine its baseband alias frequency and explain why digital processing cannot identify the original frequency from the samples alone.
2. An ADC has a ±1.25 V range and 12 bits. Calculate its quantization step and ideal full-scale sinusoidal quantization SNR.
3. Design a sampling-chain specification for a sensor whose useful band is 0–25 kHz and whose interference begins at 32 kHz. Justify the sampling rate, anti-alias transition band, quantizer range, and validation tests.

Multiple Choice Questions with Answers

Question 1

A band-limited signal has highest frequency 15 kHz. Which sampling frequency satisfies the ideal low-pass sampling theorem?

A. 20 kS/s

B. 25 kS/s

C. 30 kS/s

D. 40 kS/s

Correct Answer: D. 40 kS/s

Question 2

A 9 kHz sinusoid is sampled at 12 kS/s. Its baseband alias frequency is:

A. 1 kHz

B. 3 kHz

C. 9 kHz

D. 21 kHz

Correct Answer: B. 3 kHz

Question 3

For an ideal uniform N-bit ADC with a full-scale sinusoidal input, the approximate quantization SNR is:

A. 3.01N dB

B. 6.02N + 1.76 dB

C. 20N dB

D. 10 log₁₀(N) dB

Correct Answer: B. 6.02N + 1.76 dB

Question 4

A 10-bit ADC has a 2 V full-scale peak-to-peak range. Its quantization step is approximately:

A. 0.195 mV

B. 1.953 mV

C. 19.53 mV

D. 195.3 mV

Correct Answer: B. 1.953 mV

Question 5

Which operation must precede decimation to prevent aliasing?

A. High-pass filtering above the new Nyquist frequency

B. Low-pass filtering below the new Nyquist frequency

C. Amplitude clipping

D. Binary encoding

Correct Answer: B. Low-pass filtering below the new Nyquist frequency

Question 6

The principal purpose of an analogue anti-aliasing filter is to:

A. Increase ADC bit depth

B. Remove quantization error after conversion

C. Attenuate out-of-band energy before sampling

D. Generate DAC spectral images

Correct Answer: C. Attenuate out-of-band energy before sampling

Question 7

Under the ideal uniform quantization-noise model, the quantization-error variance is:

A. ∆/12

B. ∆²/12

C. 12∆²

D. ∆²/2

Correct Answer: B. ∆²/12

Question 8

A zero-order hold in a DAC primarily:

A. Holds each output sample for one sampling interval

B. Eliminates all spectral images without filtering

C. Raises the ADC resolution

D. Prevents input aliasing

Correct Answer: A. Holds each output sample for one sampling interval

Question 9

For small timing jitter, jitter-induced error becomes more severe when:

A. Input frequency decreases

B. Input frequency increases

C. Quantizer levels increase only

D. The signal is DC

Correct Answer: B. Input frequency increases

Question 10

Dither is commonly introduced to:

A. Increase the analogue input bandwidth

B. Decorrelate quantization error from the input

C. Guarantee zero quantization noise

D. Replace the reconstruction filter

Correct Answer: B. Decorrelate quantization error from the input