2. Sampling, Reconstruction and Quantization

Introduction and Context

Digital signal processing begins with a physical reality: most signals of interest are continuous in time, whereas digital systems operate on finite sequences of numbers. Speech, vibration, temperature, biomedical potentials, and communication waveforms are first represented as discrete-time samples and then as finite-precision digital codes. The conversion is not merely a change of notation. It imposes conditions on bandwidth, time resolution, amplitude resolution, noise, hardware complexity, storage, and system cost.

This chapter examines the signal-conversion chain that connects the analog and digital domains. Sampling converts a continuous-time waveform into a sequence indexed by integers. Reconstruction attempts to recover a continuous-time waveform from that sequence. Quantization maps each sample amplitude to one of a finite number of permitted levels, introducing error. Together, these operations form the conceptual and engineering basis of analog-to-digital converters (ADCs), digital-to-analog converters (DACs), data acquisition systems, software-defined radios, digital audio equipment, and instrumentation.

Sampling theory provides exact conditions under which a band-limited signal can be recovered without loss. In practical systems, however, signals are not perfectly band-limited, filters are nonideal, clocks have jitter, and converters have finite resolution. Consequently, a sound design requires more than memorizing the Nyquist rate. It requires analysis of aliasing, anti-alias and reconstruction filters, quantization noise, oversampling, dithering, dynamic range, signal-to-noise ratio, and converter specifications.

Learning Objectives

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

1. Distinguish continuous-time, discrete-time, and digital signals.

2. Derive and interpret the sampling theorem in both time and frequency domains.

3. Analyze aliasing and determine suitable sampling rates and filters.

4. Explain ideal and practical reconstruction using interpolation.

5. Model uniform quantization and calculate quantization error and noise power.

6. Relate converter resolution, SNR, ENOB, dynamic range, and sampling rate.

7. Explain the roles of oversampling, noise shaping, and dithering.

8. Select ADC/DAC architectures and specifications for engineering applications.

9. Identify practical errors, limitations, and professional design considerations.

From a Continuous Waveform to a Digital Representation

Let an analog signal be represented by x(t), where t is continuous. Ideal uniform sampling produces

x_s[n] = x(nT_s), n = 0, ±1, ±2, …

where T_s is the sampling interval and f_s = 1/T_s is the sampling frequency. The resulting sequence x_s[n] is discrete in time but may still have continuous-valued amplitudes. It becomes a digital signal only after each sample is quantized and encoded into a finite word.

An ideal impulse-train representation of sampling is

x_s(t) = x(t) Σ δ(t − nT_s).

Using the modulation property of the Fourier transform, the spectrum of the sampled signal is

X_s(f) = (1/T_s) Σ X(f − kf_s),

where X(f) is the spectrum of x(t), and k is any integer. Thus sampling creates periodic replicas of the original spectrum, spaced by f_s. This frequency-domain interpretation is central to sampling theory.

If the replicas remain separate, the original spectrum can be isolated with a low-pass filter. If they overlap, information is irreversibly mixed. That overlap is aliasing. No subsequent digital processing can generally determine which original frequency components produced an aliased component.

The Sampling Theorem and Its Interpretation

The classical sampling theorem states that a continuous-time signal band-limited to B hertz can be reconstructed exactly from uniform samples if

f_s > 2B.

The quantity 2B is the Nyquist rate for the signal. The term Nyquist frequency refers instead to half the sampling frequency, f_s/2. These terms are related but not interchangeable. If f_s = 20 kHz, the Nyquist frequency is 10 kHz; a signal whose highest frequency is 10 kHz has a Nyquist rate of 20 kHz.

The strict inequality is important in theory because a spectral component exactly at a replica boundary cannot be separated with a realizable filter of finite transition width. In engineering practice, the sampling frequency is selected higher than the theoretical minimum to provide a guard band for the anti-alias filter.

For a band-limited signal, ideal reconstruction is described by the sinc interpolation formula:

x(t) = Σ x[n] sinc((t − nT_s)/T_s),

where sinc(u) = sin(πu)/(πu). Each sample is associated with a shifted sinc function, and the weighted sum reproduces the original waveform. This result is exact under the assumptions of band limitation, uniform sampling, noiseless measurements, and infinite-duration ideal filtering. Real systems approximate these conditions.

Bandpass Sampling

The requirement f_s > 2B does not necessarily mean that f_s must exceed twice the highest frequency present. If a signal occupies a narrow band between f_L and f_H, with bandwidth B = f_H − f_L, carefully chosen undersampling can translate that band to a lower-frequency alias without overlap. This is called bandpass sampling or intentional undersampling.

Bandpass sampling is useful in radio receivers and high-frequency instrumentation, but it demands strict control of spectral occupancy. The sampling rate must ensure that translated spectral bands do not overlap, and the analog input must be filtered to remove unwanted bands. It is therefore a designed frequency-translation technique, not an excuse to sample arbitrary broadband signals below twice their highest frequency.

Aliasing: Irreversible Spectral Ambiguity

A sinusoid illustrates aliasing clearly. For

x(t) = A cos(2πf_0t + φ),

the samples are

x[n] = A cos(2πf_0nT_s + φ).

Frequencies differing by integer multiples of f_s produce the same sampled sequence. The apparent frequency can therefore be found by folding the input frequency into the interval from 0 to f_s/2. For example, sampling a 7 kHz sinusoid at 10 kHz produces samples indistinguishable from a 3 kHz sinusoid, because 7 kHz = 10 kHz − 3 kHz.

Aliasing may arise from an input frequency above f_s/2, from an inadequately filtered interferer, or from sampling-clock modulation. It is especially dangerous in measurement systems because it can create plausible but false low-frequency information. A vibration sensor may report a low-frequency mechanical oscillation that is actually a high-frequency machine fault; an imaging system may display spatial patterns that do not exist in the scene.

The anti-alias filter is an analog low-pass or band-pass filter placed before the ADC. Its purpose is not to make the desired signal digital; it limits out-of-band energy before sampling. A practical design divides the spectrum into a passband, transition band, and stopband. The sampling rate, permitted distortion, and required attenuation determine the filter order and implementation cost.

Example: Selecting a Sampling Rate

Suppose an instrumentation channel must preserve frequencies up to 18 kHz, while frequencies above 24 kHz must be attenuated by at least 80 dB. Sampling at 36 kHz leaves no transition band and is impractical for a finite-order anti-alias filter. A rate such as 60 kHz provides a transition region from 18 to 30 kHz, making the analog filter more realizable. The selected rate must also account for clock tolerance, component variation, and possible out-of-band interference.

Reconstruction and Interpolation

Reconstruction is the reverse conceptual operation of sampling. An ideal DAC may be modeled as an impulse generator followed by an ideal low-pass filter. The filter removes spectral images centered at multiples of f_s and retains the baseband spectrum. In time-domain language, reconstruction is interpolation between samples.

The zero-order hold, widely used in practical DACs, holds each digital sample constant until the next sample arrives. Its frequency response is

H_ZOH(f) = T_s sinc(fT_s)e^{-jπfT_s}.

The sinc factor causes amplitude droop as frequency approaches the Nyquist frequency, and the exponential term represents a delay. A reconstruction filter is commonly placed after the DAC to attenuate spectral images and smooth the staircase-like output.

Digital interpolation can precede the DAC. An interpolator inserts zeros between samples and applies a digital low-pass filter. This increases the effective sample rate and moves spectral images farther from the desired band. The subsequent analog reconstruction filter can then have a wider transition band and lower order. Oversampling DACs exploit this principle extensively.

Practical reconstruction filters may be analog Butterworth, Chebyshev, elliptic, Bessel, or active-RC designs, depending on the trade-off between amplitude flatness, phase response, transition sharpness, and implementation complexity. In precision measurement, phase linearity may be more important than maximum stopband attenuation; in communication systems, adjacent-channel rejection may dominate.

Quantization: Amplitude Discretization

Sampling discretizes time, but quantization discretizes amplitude. A uniform quantizer with N bits has L = 2^N representation levels. If the input range is from −V_FS/2 to +V_FS/2, the nominal step size is

Δ = V_FS/L.

A sample is replaced by the nearest permitted level in a mid-tread or mid-rise quantizer, depending on how the zero region is represented. The quantization error is

e_q[n] = x[n] − x_q[n].

For an ideal rounding quantizer, the error lies approximately between −Δ/2 and +Δ/2. Under the common assumptions that the signal exercises many quantizer levels, the error is sufficiently uncorrelated with the signal, and its probability density is approximately uniform, the mean-square quantization noise is

σ_q^2 = Δ^2/12.

This is a model, not a universal law. For low-level, periodic, or highly correlated inputs, quantization error may be deterministic and strongly related to the signal. It can then appear as spurious tones rather than white noise.

For a full-scale sinusoid applied to an ideal N-bit converter, the approximate signal-to-quantization-noise ratio is

SNR_q ≈ 6.02N + 1.76 dB.

The formula assumes that the sinusoid uses the available range without clipping and that quantization noise is spread across the Nyquist band. A 12-bit ideal converter therefore has approximately 74 dB SNR under these conditions. The formula should not be applied blindly to real converters, whose thermal noise, distortion, clock jitter, reference noise, and nonlinearity reduce performance.

Dynamic Range and Effective Number of Bits

Dynamic range is the ratio, usually expressed in decibels, between a specified maximum usable signal and the noise floor. Effective number of bits, or ENOB, summarizes the combined impact of noise and distortion. If SINAD is the signal-to-noise-and-distortion ratio in decibels,

ENOB = (SINAD − 1.76)/6.02.

Resolution is a nominal architectural property; ENOB is a measured performance property. A 16-bit ADC may have an ENOB of only 13.5 bits at a particular input frequency. Conversely, oversampling and averaging may improve in-band resolution without changing the converter's nominal code width.

Oversampling, Noise Shaping, and Dithering

Oversampling means sampling substantially faster than the minimum required rate. If quantization noise is approximately white over the Nyquist band, increasing f_s spreads the same total noise power over a wider frequency range. A subsequent digital filter can retain the signal band and reject part of the noise. If the oversampling ratio is doubled, the in-band quantization noise power ideally decreases by a factor of two, corresponding to approximately 3 dB improvement in SNR for a fixed signal bandwidth.

Noise shaping goes further by using feedback to move quantization noise away from the signal band. Sigma-delta converters employ oversampling and noise shaping, followed by digital filtering and decimation. Their high resolution at low and moderate bandwidths makes them common in audio, precision instrumentation, and sensor interfaces. Their trade-offs include latency, out-of-band noise, stability considerations, and dependence on digital filtering.

Dithering adds a controlled noise signal before quantization. Although this may appear undesirable, properly designed dither decorrelates quantization error from the input and suppresses deterministic distortion or idle tones. Dither does not eliminate noise; it exchanges correlated spurious behavior for a more noise-like and predictable error. In professional audio, imaging, and precision measurement, the spectral distribution and amplitude of dither must be chosen carefully.

ADC and DAC Architectures in Engineering Practice

The converter architecture should be matched to bandwidth, resolution, latency, power, cost, and environmental requirements.

A flash ADC uses a bank of comparators and can achieve very high conversion speed, but its hardware grows rapidly with resolution because approximately 2^N − 1 comparators are required. It is suitable for high-speed applications where moderate resolution is acceptable.

A successive-approximation-register (SAR) ADC uses a sample-and-hold circuit, comparator, and internal DAC in a binary-search procedure. SAR converters offer a strong balance of resolution, speed, power, and latency and are widely used in control, data acquisition, and industrial instrumentation.

A pipeline ADC divides conversion across stages. It provides high throughput and moderate-to-high resolution, usually with several clock cycles of latency. It is common in communications, imaging, and measurement systems.

A sigma-delta ADC oversamples, shapes quantization noise, and digitally filters the result. It provides excellent in-band resolution but generally has greater latency and limited bandwidth compared with high-speed pipeline designs.

DACs may use binary-weighted current or voltage sources, R-2R resistor networks, segmented architectures, or pulse-density and sigma-delta techniques. Important DAC errors include offset, gain error, differential nonlinearity, integral nonlinearity, glitch impulse, settling time, and output compliance.

Specifications That Matter in Converter Selection

A professional specification review should distinguish static and dynamic performance. Offset error is the output error at zero input after ideal scaling. Gain error is the slope deviation after offset is removed. Integral nonlinearity measures deviation from an ideal transfer curve, while differential nonlinearity measures code-width variation. Missing codes and nonmonotonic behavior are particularly serious in control applications.

Dynamic specifications include SNR, SINAD, total harmonic distortion, spurious-free dynamic range, aperture uncertainty, and effective number of bits. The input frequency matters: a converter's ENOB may decline at higher frequencies because of front-end settling limits and aperture jitter.

For a sinusoidal input of frequency f_in, RMS timing-jitter noise has the approximate limitation

SNR_jitter = −20 log10(2π f_in σ_t),

where σ_t is the RMS sampling-clock jitter. Thus the same clock may be adequate at low frequency but inadequate for a high-frequency communication signal. Reference stability, grounding, electromagnetic interference, power-supply rejection, input-drive circuitry, and thermal management are often as important as the converter data sheet.

Applied Case Study: Digitizing a Vibration-Monitoring Channel

Background

Consider an illustrative condition-monitoring system for an industrial motor. An accelerometer produces useful information from 0.5 kHz to 22 kHz. A known switching component can generate interference near 31 kHz, and occasional broadband electrical transients occur above 100 kHz.

Problem

The system must identify changes in bearing vibration without confusing high-frequency interference with genuine low-frequency faults. The designer must select the sampling rate, analog filtering, ADC resolution, and digital processing strategy.

Analysis

Sampling at 44 kHz would satisfy the theoretical Nyquist rate for 22 kHz, but it would leave no practical transition band before the 22 kHz Nyquist frequency. The 31 kHz interferer would also alias into the measured band if not strongly attenuated. A higher rate, such as 96 kHz, places the Nyquist frequency at 48 kHz and provides a transition region between the 22 kHz passband and the 48 kHz stopband boundary.

The anti-alias filter should pass the required vibration band with acceptable amplitude and phase error, attenuate the 31 kHz component sufficiently, and suppress higher-frequency transients. The ADC resolution is selected from the smallest vibration amplitude to be detected, while gain staging ensures that large transients do not cause persistent clipping. A SAR converter may be appropriate when moderate latency and wide bandwidth are required. If the useful band were narrower and resolution more important than latency, a sigma-delta converter could be considered.

Approach and Solution

The proposed signal chain uses an accelerometer interface, analog gain and protection, a low-pass anti-alias filter, a 96 kS/s ADC, digital band-pass filtering, and spectral analysis. Calibration is performed using known reference vibration levels. The system records converter overloads and monitors the noise floor so that apparent changes are not attributed automatically to mechanical faults.

Outcome

The wider transition band reduces the required analog filter order, while digital filtering provides selective analysis after conversion. The 31 kHz interference is prevented from masquerading as a lower-frequency vibration component. The design also retains diagnostic information above the principal analysis band, should future algorithms require it.

Learning Points

The case illustrates that the sampling rate cannot be selected from the desired passband alone. Filter realizability, interference location, converter performance, and future processing requirements must be considered together. It also demonstrates why aliasing prevention must occur before the ADC; digital filtering cannot recover information already corrupted by aliasing.

Limitations, Risks, and Common Design Errors

Several recurring mistakes undermine otherwise sophisticated DSP systems. Confusing the Nyquist rate with the Nyquist frequency leads to incorrect specifications. Sampling exactly at twice the highest frequency ignores the need for a transition band. Omitting an anti-alias filter assumes that the environment is band-limited, which is rarely justified. Treating quantization error as always white noise can hide tones and idle patterns. Using nominal resolution as a substitute for ENOB ignores analog noise and distortion.

Clipping is another distinct failure: unlike quantization noise, clipping produces severe nonlinear distortion and harmonics. Poor clock design introduces jitter; poor layout couples digital switching into sensitive analog inputs; and inadequate reference decoupling changes converter scale and linearity. In DAC systems, image components and zero-order-hold droop may be neglected if the output is inspected only at low frequency.

Sampling systems also have a measurement-governance dimension. Calibration intervals, traceability, overload indicators, timestamp integrity, and preservation of raw data are important where measurements support safety, maintenance, medical decisions, or regulatory reporting. A falsely precise digital value is not an accurate measurement.

Best-Practice Design Workflow

A robust design commonly follows this sequence:

1. Define the required signal band, amplitude range, allowable distortion, latency, and noise floor.

2. Identify known interferers and plausible out-of-band energy.

3. Select the sampling frequency with a practical transition band rather than using only the theoretical minimum.

4. Design the analog anti-alias filter and verify attenuation under component tolerances.

5. Select ADC resolution and architecture using ENOB, SNR, bandwidth, power, and latency—not nominal bits alone.

6. Check clock-jitter limits at the highest input frequency.

7. Plan digital filtering, decimation, interpolation, and reconstruction requirements.

8. Test with swept-frequency, multitone, low-level, full-scale, and out-of-band signals.

9. Verify calibration, overload behavior, electromagnetic compatibility, and long-term drift.

Current Developments and Emerging Directions

Modern data converters increasingly combine analog conversion with substantial digital correction. Background calibration can compensate gain, offset, timing skew, and some nonlinearities while the system operates. Time-interleaved ADCs increase aggregate sampling rate by using several converters in parallel, but mismatches between channels create spurious tones that require calibration.

Converter design is also being influenced by edge computing and intelligent sensing. Local digital filtering and feature extraction can reduce data transport, but the front end must still preserve the information needed by algorithms. Machine-learning systems do not remove sampling requirements: if an important feature is aliased or clipped before inference, later computation cannot reconstruct it reliably. In high-speed communications, digitally assisted analog filtering, direct RF sampling, and advanced clock recovery continue to extend usable bandwidth, while power efficiency and thermal limits remain central constraints.

Chapter Summary

Sampling converts a continuous-time signal into a discrete-time sequence, while quantization converts continuous amplitudes into finite numerical codes. Uniform sampling creates periodic spectral replicas; exact reconstruction is possible when the signal is sufficiently band-limited and the sampling frequency exceeds twice its bandwidth. Aliasing occurs when spectral replicas overlap and cannot be removed after sampling, making anti-alias filtering essential.

Reconstruction uses interpolation and low-pass filtering to remove DAC images, while the zero-order hold introduces frequency-dependent droop. Quantization introduces an error whose ideal noise power is Δ²/12 under suitable assumptions, leading to the approximate full-scale sinusoidal SNR of 6.02N + 1.76 dB. Real converter performance is described more accurately by SNR, SINAD, ENOB, SFDR, linearity, jitter, and dynamic range.

Oversampling spreads quantization noise, noise shaping moves it away from the signal band, and dithering reduces correlation and deterministic artifacts. ADC/DAC selection requires integrated consideration of bandwidth, filters, resolution, latency, clock quality, interference, power, calibration, and application risk. The central engineering principle is that analog limitations must be addressed before information enters the digital domain.

Key Terms

Sampling: The process of measuring a continuous-time signal at discrete time instants.

Sampling frequency: The number of samples acquired per second, denoted f_s.

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

Nyquist frequency: Half the sampling frequency, f_s/2.

Aliasing: Irreversible spectral ambiguity caused by overlapping sampled-spectrum replicas.

Anti-alias filter: An analog filter that limits input bandwidth before an ADC.

Reconstruction: Recovery of a continuous-time waveform from discrete samples.

Interpolation: Estimation of values between known samples.

Zero-order hold: A DAC output mechanism that holds each sample constant until the next sample.

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

Quantization step: The voltage interval between adjacent quantization levels.

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

Oversampling: Sampling at a rate substantially higher than the minimum required rate.

Noise shaping: Spectral redistribution of quantization noise, usually through feedback.

Dithering: Addition of controlled noise to reduce correlation between signal and quantization error.

ENOB: Effective number of bits inferred from measured SINAD.

SINAD: Ratio of signal power to combined noise and distortion power.

Aperture jitter: Sampling-time uncertainty that produces input-dependent error.

Differential nonlinearity: Variation of actual code width from the ideal code width.

Spurious-free dynamic range: Difference between the desired signal and the largest undesired spectral component.

Review Questions

Short Answer Questions

1. Distinguish the Nyquist rate from the Nyquist frequency.

2. Why is an anti-alias filter placed before an ADC?

3. State the ideal quantization-noise variance for a uniform rounding quantizer.

4. What is the function of a reconstruction filter after a DAC?

5. Define ENOB and explain why it may be lower than nominal resolution.

Long Answer Questions

1. Derive the frequency-domain representation of an impulse-sampled signal and explain the condition for ideal reconstruction.

2. Explain aliasing using a sinusoidal example and discuss how practical filters influence sampling-rate selection.

3. Describe the operation and trade-offs of flash, SAR, pipeline, and sigma-delta ADCs.

4. Explain quantization noise, the 6.02N + 1.76 dB relationship, and the conditions under which it is valid.

5. Discuss oversampling, noise shaping, and dithering as methods for improving practical converter performance.

Analytical / Application Questions

1. A 15 kHz low-pass signal is sampled at 32 kS/s. Determine whether ideal sampling theory permits recovery and identify practical filter-design concerns.

2. A 10-bit ADC has a 4.096 V full-scale range. Calculate its ideal step size and approximate full-scale sinusoidal quantization SNR.

3. A vibration system must preserve 20 kHz but contains a strong 28 kHz interferer. Propose a sampling rate and filtering strategy, and justify the choice in terms of aliasing and transition-band width.

22. Multiple Choice Questions with Answers

Question 1

A signal is sampled at 20 kS/s. Which frequency is its Nyquist frequency?

A. 5 kHz

B. 10 kHz

C. 20 kHz

D. 40 kHz

Correct Answer: B. 10 kHz

Question 2

A 7 kHz sinusoid sampled at 10 kS/s appears as which baseband frequency?

A. 1 kHz

B. 3 kHz

C. 7 kHz

D. 17 kHz

Correct Answer: B. 3 kHz

Question 3

What is the principal purpose of an anti-alias filter?

A. To increase ADC resolution

B. To remove quantization noise after conversion

C. To limit out-of-band energy before sampling

D. To reconstruct the continuous-time output

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

Question 4

For an ideal 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. N/6.02 dB

Correct Answer: B. 6.02N + 1.76 dB

Question 5

A 12-bit ADC has an ideal full-scale sinusoidal quantization SNR closest to

A. 37 dB

B. 62 dB

C. 74 dB

D. 96 dB

Correct Answer: C. 74 dB

Question 6

Which statement best describes oversampling?

A. It eliminates all analog filtering requirements

B. It spreads quantization noise over a wider frequency range

C. It reduces the number of quantization levels

D. It guarantees zero clock jitter

Correct Answer: B. It spreads quantization noise over a wider frequency range

Question 7

Dithering is introduced primarily to

A. Increase the sampling frequency

B. Eliminate the need for a DAC

C. Decorrelate quantization error from the input

D. Produce spectral images intentionally

Correct Answer: C. Decorrelate quantization error from the input

Question 8

If a converter has SINAD = 74 dB, its approximate ENOB is

A. 8 bits

B. 10 bits

C. 12 bits

D. 16 bits

Correct Answer: C. 12 bits

Question 9

Which ADC architecture generally uses a binary-search procedure?

A. Flash

B. SAR

C. Pipeline only

D. Sigma-delta only

Correct Answer: B. SAR

Question 10

Why can sampling-clock jitter be especially limiting at high input frequencies?

A. Timing uncertainty produces larger voltage error when the waveform changes rapidly

B. Jitter increases the number of ADC codes

C. Jitter removes all spectral images

D. Timing uncertainty affects only DC signals

Correct Answer: A. Timing uncertainty produces larger voltage error when the waveform changes rapidly