Sampling, Reconstruction, and Aliasing

Introduction and Context

Sampling is the process of representing a continuous-time signal by values measured at discrete instants. It is the fundamental interface between physical phenomena—sound, temperature, vibration, radio waves—and digital computation. Reconstruction is the inverse operation: producing a continuous-time waveform, or an equivalent continuous-time representation, from samples. Aliasing is the loss of uniqueness that occurs when sampling is performed below the rate required to distinguish the signal’s frequency components.

Sampling theory is central to analog-to-digital conversion, digital communications, software-defined radio, medical imaging, audio engineering, control, instrumentation, and machine learning hardware. At postgraduate level, it is important to distinguish the ideal mathematical sampling model from practical sample-and-hold circuits, finite-bandwidth systems, clock uncertainty, quantization, and intentional undersampling.

Learning Objectives

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

1. Explain ideal and practical sampling in time and frequency domains.
2. Derive and apply the Nyquist–Shannon sampling theorem.
3. Determine spectral replicas and identify aliasing conditions.
4. Design reconstruction filters and interpret zero-order-hold behavior.
5. Analyze bandpass sampling and frequency translation.
6. Quantify the effects of aperture uncertainty, timing jitter, and finite record length.
7. Select suitable sampling architectures for engineering applications.

1. Continuous-Time Signals and Their Spectra

Let x(t) denote a continuous-time signal, where t is measured in seconds. Its Fourier transform is defined, under the usual integrability or generalized-function conditions, by

Here f is frequency in hertz and j = √−1. The inverse transform is

A signal is strictly band-limited to B hertz if X(f) = 0 for |f| > B. Band limitation is an idealization, because physical signals generally have spectral tails, but engineering systems can often impose a sufficiently strong effective bandwidth limit.

The highest frequency present in a real low-pass signal is commonly called its bandwidth edge or highest occupied frequency. It is not necessarily the same as the signal’s average frequency, carrier frequency, or information rate.

2. Ideal Impulse-Train Sampling

Uniform sampling takes values at times t = nT_s, where n is an integer and T_s is the sampling interval. The sampling frequency is

with units of samples per second or hertz. The ideal sampling function is an impulse train:

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

This model records ideal instantaneous amplitudes and does not include quantization, aperture duration, or circuit loading.

The Fourier transform of the impulse train is another impulse train:

Since multiplication in time corresponds to convolution in frequency,

Thus, sampling replicates the original spectrum around every integer multiple of f_s. The factor f_s results from the transform convention and preserves dimensional consistency.

Spectral interpretation

If X(f) occupies −B ≤ f ≤ B, each replica occupies

for integer k. Perfect recovery is possible when neighboring replicas do not overlap. For a low-pass signal, the sufficient condition is

The strict inequality is preferred because equality causes replicas to meet at a boundary and is highly sensitive to phase, filter roll-off, and nonidealities.

3. Nyquist–Shannon Sampling Theorem

Theorem. If x(t) is band-limited to B hertz and is sampled uniformly at a rate f_s > 2B, then x(t) is uniquely determined by its samples x(nT_s), provided ideal samples and ideal reconstruction conditions hold.

The minimum rate 2B is called the Nyquist rate of the signal. The interval T_s/2 is sometimes called the Nyquist interval. The phrase “Nyquist frequency” usually means f_s/2, the highest frequency represented without aliasing in a sampled system. These terms must not be confused.

The ideal interpolation formula is

where

This expression assumes a strictly band-limited signal, infinite-duration data, exact samples, and an ideal brick-wall reconstruction filter. The sinc functions extend infinitely in time, so the formula is theoretically exact but not directly realizable as a finite-latency filter.

Derivation through spectral selection

When f_s > 2B, the central replica X(f) is separated from its neighbors. An ideal low-pass filter with passband |f| ≤ B can select the central replica. If the filter has gain T_s in the passband, then

and inverse transformation gives x_r(t) = x(t). The reconstruction filter must reject all spectral replicas while retaining the desired baseband.

4. Aliasing

Aliasing occurs when distinct continuous-time frequencies produce the same sample sequence. For a complex sinusoid x(t) = e^{j2πf_0t}, samples are

Frequencies f_0 and f_0 + kf_s generate identical samples for every integer k because

Therefore frequency is observable only modulo f_s. For real sinusoids, positive and negative frequency symmetry further produces the familiar folding around f_s/2.

The aliased frequency can be found by choosing an integer k such that

For a signal with multiple components, each component must be examined separately. Aliasing is irreversible after sampling: no digital filter can determine which original frequency produced an aliased component.

Worked Example 1: Frequency alias

A sinusoid at 9 kHz is sampled at 10 kHz. Choose k = 1:

The 9 kHz sinusoid therefore appears as a 1 kHz component in the sampled data. A 1 kHz analog sinusoid sampled at 10 kHz produces the same samples, apart from possible phase conventions.

Worked Example 2: Sampling-rate selection

An instrumentation signal occupies 0–18 kHz. The theoretical Nyquist rate is 36 kHz. A practical design might select 48 kS/s or higher, depending on the transition band required by the analog anti-alias filter. If the analog filter must attenuate frequencies above 20 kHz and the converter operates at 48 kS/s, the Nyquist frequency is 24 kHz, leaving a 4 kHz transition band. A converter operating at 40 kS/s would leave only a 0 kHz margin between 20 kHz and the Nyquist frequency and would be unsuitable for a realistic transition band.

5. Anti-Aliasing and Reconstruction Filters

An anti-aliasing filter is an analog low-pass or band-pass filter placed before an ADC. Its purpose is to attenuate out-of-band energy before sampling. It cannot remove energy already aliased into the desired band.

A practical filter is specified by:

• passband edge and allowable passband ripple;
• stopband edge and required attenuation;
• transition width;
• phase or group-delay requirements;
• implementation order and power consumption.

For a low-pass signal, let f_p be the maximum passband frequency, f_stop the frequency at which the required stopband attenuation must begin, and f_samp the sampling frequency. A feasible design requires

The reconstruction filter after a DAC performs the complementary operation: it suppresses images around multiples of f_s. In practice, oversampling relaxes both anti-alias and reconstruction filter requirements by increasing the transition band.

6. Sample-and-Hold and Zero-Order Hold

Real ADCs measure over a finite aperture or use a track-and-hold circuit. A DAC commonly holds each sample constant until the next sample. The zero-order-hold impulse response is

Its frequency response is

The magnitude is

The factor T_s reflects the hold’s time dimension. If normalized gain is used, the response is often written as sinc(fT_s). The response causes passband droop and phase delay. A reconstruction filter may include equalization, but equalization cannot recover frequencies removed by a zero of the hold response.

7. Bandpass Sampling and Undersampling

The condition f_s > 2B is sufficient for low-pass sampling, but it is not the only possible strategy. Suppose a signal occupies a band f_L ≤ f ≤ f_H, with bandwidth B = f_H − f_L. Sampling can be valid even when f_s is less than 2f_H if spectral replicas do not overlap.

For a real bandpass signal, a candidate sampling rate must satisfy the no-overlap constraints for the translated bands. A commonly used practical criterion is

for a positive integer m, with an interval that must actually be nonempty and must be checked against guard bands and filter roll-off. This expression is a design aid, not a substitute for plotting spectral replicas. The translated band may be inverted in frequency, which is acceptable if the receiver accounts for spectral inversion.

Worked Example 3: Bandpass sampling

A band-limited RF signal occupies 70–80 MHz, so B = 10 MHz. For m = 2, the candidate interval is

which gives 80 MHz ≤ f_s ≤ 140 MHz. A rate of 100 MS/s is therefore a candidate. The original band translates to an intermediate band rather than baseband, and a detailed replica check is required. The analog input filter remains essential: signals outside 70–80 MHz could overlap the desired translated band.

Bandpass sampling is widely used in software-defined radios, oscilloscopes, and direct RF receivers. It is deliberate undersampling, whereas aliasing is unwanted ambiguity. The mathematical mechanism is the same; engineering intent and filtering distinguish the two.

8. Sampling of Complex Signals and I/Q Systems

A complex analytic signal can be represented with one-sided spectrum, reducing the apparent sampling burden relative to a real passband signal. In-phase and quadrature components are commonly written

where I(t) and Q(t) are real signals. An ideal complex downconverter can translate a selected RF band to complex baseband. The required sampling rate is governed by the occupied complex bandwidth and the implementation’s image-rejection performance, not simply by the RF carrier frequency.

In practical I/Q systems, gain mismatch, phase imbalance, DC offsets, oscillator phase noise, and finite analog filtering create image leakage and spurious responses. Calibration and digital compensation are therefore part of the sampling architecture.

9. Nonuniform Sampling and Reconstruction

Uniform sampling leads to simple spectral periodicity. Nonuniform sampling uses times t_n that are not equally spaced. It can be advantageous in event-driven sensing, astronomy, biomedical acquisition, and compressed sensing, but reconstruction becomes a model-dependent inverse problem.

For nonuniform samples, ordinary DFT assumptions do not apply directly. Algorithms such as nonuniform FFT methods, generalized interpolation, and sparse reconstruction may be used. Uniqueness depends on sampling geometry, signal class, noise, and prior information. A signal that is recoverable under a sparse model may not be recoverable as an arbitrary band-limited waveform.

Compressed sensing is not a license to sample below the information required by every signal. It relies on sparsity or compressibility, an appropriate measurement operator, and a reconstruction algorithm. Violating the model can produce plausible but incorrect reconstructions.

10. Practical Nonidealities

Quantization

Sampling discretizes time; quantization discretizes amplitude. For an ideal uniform ADC with N bits and full-scale range V_FS, the quantization step is approximately

For an input that fully spans the quantizer and has sufficiently distributed quantization error, the quantization-noise power is approximately

This model assumes an ideal quantizer, no overload, and an error that behaves approximately as an uncorrelated uniform random variable. Quantization noise is distinct from aliasing.

Aperture jitter

If sample timing has rms uncertainty σ_t, a sinusoid of frequency f_in experiences an approximate jitter-limited signal-to-noise ratio

This approximation assumes small timing errors and random jitter. It shows why high-frequency inputs demand exceptionally stable clocks. For example, f_in = 100 MHz and σ_t = 1 ps yield

and SNR_jitter ≈ 64 dB.

Clock phase noise and drift

Random jitter produces noise-like degradation, while deterministic periodic timing modulation produces spurs and sidebands. Long-term clock drift changes the effective sampling rate and can cause frequency-scale errors. In coherent systems, clock synchronization and reference distribution are therefore signal-integrity issues, not merely timing details.

Finite observation and leakage

A finite record of N samples is equivalent to multiplying the infinite sequence by a rectangular window. In frequency, this convolves the spectrum with the window spectrum, producing leakage. Window functions such as Hann, Blackman, or flat-top windows trade frequency resolution, amplitude accuracy, and sidelobe suppression. Leakage is not aliasing, although both can appear as unwanted spectral components.

11. Reconstruction in Digital Signal Processing

Digital reconstruction may mean either recovering an analog waveform with a DAC or estimating signal values between samples. Ideal sinc interpolation is conceptually exact under theorem assumptions. Practical systems use finite impulse-response (FIR), infinite impulse-response (IIR), polyphase, or windowed-sinc filters.

An interpolation-by-L system inserts L−1 zeros between samples and applies a low-pass filter. If the original sequence has sampling frequency f_s, the interpolated sequence has frequency Lf_s. The interpolation filter removes spectral images created by zero insertion and normally has passband gain L under the standard multirate convention.

Decimation by M first requires low-pass filtering to prevent aliasing, then retains every Mth sample. If the original rate is f_s, the output rate is f_s/M. The pre-decimation cutoff must be no greater than the new Nyquist frequency f_s/(2M), subject to transition-band design.

A rational sampling-rate converter with factors L and M performs interpolation by L followed by decimation by M, giving

The anti-alias filter is placed before the decimation operation, and efficient polyphase structures avoid computing samples that will later be discarded.

12. A Design Workflow

A robust sampling design can follow these steps:

1. Define the information-bearing analog band, including allowable out-of-band interference.
2. Determine whether the signal is low-pass, band-pass, real, or complex.
3. Choose a sampling rate with margin rather than operating exactly at the theoretical limit.
4. Specify analog anti-alias filtering from passband, stopband, attenuation, and transition width.
5. Evaluate ADC resolution, full-scale range, noise, distortion, aperture jitter, and clock phase noise.
6. Model sample-and-hold and DAC reconstruction behavior.
7. Select digital decimation, interpolation, and equalization filters.
8. Verify spectral replicas, spurious responses, and worst-case interference.
9. Validate the design using measured spectra, coherent test tones, and out-of-band stress tests.

Oversampling can reduce analog filter complexity and improve noise-shaping opportunities, but it increases data rate, power, memory, and processing load. The best rate is therefore a system-level trade-off rather than simply the largest available rate.

13. Illustrative Case Study: Vibration Monitoring of a Motor

Background

Consider an industrial motor monitored for bearing faults. The useful vibration band is 0–18 kHz. Strong mechanical and electrical interference may occur near 30–40 kHz.

Problem

A proposed ADC operates at 32 kS/s. Its Nyquist frequency is 16 kHz, below the desired 18 kHz band. In addition, the 30–40 kHz interference can fold into the measurement band.

Analysis

The proposed rate violates the low-pass sampling condition because 32 kS/s < 2(18 kHz) = 36 kS/s. A 17 kHz component, for example, aliases to |17 − 32| = 15 kHz. A 35 kHz interferer aliases to |35 − 32| = 3 kHz. The latter could be incorrectly interpreted as a low-frequency mechanical feature.

Approach and solution

The engineering team selects a 64 kS/s ADC, whose Nyquist frequency is 32 kHz, and designs an analog anti-alias filter that preserves 0–18 kHz while strongly attenuating interference beginning near 28 kHz. The remaining digital data are decimated to 16 kS/s only after a digital low-pass filter limits the band to below 8 kHz for a separate low-frequency diagnostic channel. The high-frequency diagnostic channel is retained at 64 kS/s.

Outcome

The design provides a realistic analog transition band, prevents the 35 kHz interference from entering the original measurement, and permits separate digital channels with independently justified bandwidths. The cost is increased ADC throughput and more careful clock and filter design.

Learning points

Sampling rate must be selected from the complete analog spectrum, not merely the nominal feature of interest. Filtering before sampling is essential, and later digital filtering cannot undo analog aliasing.

14. Limitations, Governance, and Contemporary Practice

Sampling decisions affect safety and compliance in medical devices, aircraft instrumentation, industrial protection, and communications interception. Requirements should document bandwidth assumptions, calibration, traceability, overload behavior, clock quality, and test evidence. In safety-critical systems, an undocumented assumption that “the signal is band-limited” is a design risk.

Modern converters use oversampling, delta–sigma modulation, time-interleaved ADCs, digital calibration, and integrated anti-alias filtering. Time-interleaved converters increase aggregate rate but can introduce channel-skew spurs and gain mismatch. Delta–sigma converters shape quantization noise out of band and rely on digital decimation filters; their apparent high resolution does not eliminate analog interference constraints.

AI-based reconstruction and denoising can estimate missing or corrupted samples, but such methods do not restore information destroyed by unrestricted aliasing. They impose learned priors and may generate plausible details unsupported by measurements. In scientific and medical applications, reconstructed content must be labeled, validated, and separated from directly observed data.

Best Practices

• State the signal bandwidth and the assumptions supporting it.
• Use an analog anti-alias filter before every ADC unless a justified integrated alternative exists.
• Do not design exactly at the Nyquist limit when practical filter transition bands are required.
• Plot spectral replicas for unusual sampling schemes, especially bandpass sampling.
• Distinguish aliasing, quantization noise, leakage, distortion, and clock-jitter effects.
• Use coherent sampling or suitable windows when measuring spectra.
• Verify worst-case interferers, overload, and clock performance experimentally.
• Treat model-based or AI reconstruction as inference, not as recovery of information that was never sampled uniquely.

Chapter Summary

Uniform sampling creates periodic replicas of a continuous-time spectrum at integer multiples of the sampling frequency. A band-limited signal can be uniquely reconstructed when the sampling frequency exceeds twice its highest frequency, under ideal sampling and filtering assumptions. Aliasing occurs when spectral replicas overlap or when distinct analog frequencies produce identical sample sequences; it cannot be removed after sampling. Anti-alias filters, reconstruction filters, sample-and-hold modeling, oversampling, and careful clock design are central to practical systems. Bandpass sampling can intentionally translate a selected RF band to a lower frequency, but it requires replica analysis and strict analog filtering. Quantization, timing jitter, finite-record leakage, and converter nonidealities are separate mechanisms that must be analyzed alongside aliasing. Modern multirate DSP, software-defined radio, vibration monitoring, and data-conversion systems all rely on these principles.

Key Terms

Sampling: Conversion of a continuous-time waveform into values at specified time instants.

Sampling interval: Time between consecutive uniform samples, denoted T_s.

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

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

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

Aliasing: Irreversible ambiguity in which different analog frequencies produce the same samples.

Spectral replica: A shifted copy of a signal spectrum created by sampling.

Anti-alias filter: Analog filter that attenuates unwanted components before sampling.

Reconstruction filter: Filter that removes spectral images after digital-to-analog conversion.

Sinc interpolation: Ideal interpolation using shifted sinc functions.

Zero-order hold: Circuit or model that holds each sample value constant for one sampling interval.

Bandpass sampling: Sampling a selected band at a rate lower than twice its highest RF frequency without overlap of replicas.

Quantization: Discretization of amplitude into finite numerical levels.

Quantization noise: Error introduced by amplitude quantization.

Aperture jitter: Uncertainty in the instant at which a sample is acquired.

Oversampling: Sampling substantially faster than the minimum theoretical rate.

Decimation: Low-pass filtering followed by reduction of sampling rate.

Interpolation: Increase of sampling rate, normally accompanied by image-suppressing filtering.

Spectral leakage: Spreading in a finite-record spectrum caused by windowing and observation truncation.

Review Questions

Short Answer Questions

1. Define the Nyquist rate and distinguish it from the Nyquist frequency.
2. Why does ideal sampling produce spectral replicas?
3. What is the function of an anti-aliasing filter?
4. Explain why aliasing cannot generally be corrected by a digital filter after sampling.
5. State two differences between quantization noise and timing jitter.

Long Answer Questions

1. Derive the spectrum of an impulse-train sampled signal and explain the no-overlap condition.
2. Explain the Nyquist–Shannon theorem and the sinc interpolation formula, including their assumptions.
3. Compare low-pass sampling, bandpass sampling, and complex I/Q sampling.
4. Discuss zero-order hold, reconstruction filtering, and passband droop in a DAC system.
5. Develop a complete sampling design procedure for an industrial measurement system.

Analytical / Application Questions

1. A signal occupies 0–22 kHz and contains an interferer at 47 kHz. Determine whether sampling at 50 kS/s is safe without an analog anti-alias filter, and calculate the apparent frequency of the interferer.
2. A 12-bit ADC has a 4.096 V full-scale range. Calculate its ideal quantization step and approximate quantization-noise variance, stating the assumptions.
3. A 2.4 GHz receiver uses a 20 MHz-wide band from 2.40 to 2.42 GHz. Propose a candidate bandpass sampling rate, verify the replica condition, and identify the analog filtering requirement.

Multiple Choice Questions with Answers

Question 1

A signal band-limited to 15 kHz is sampled uniformly. Which rate satisfies the ideal low-pass sampling theorem?

A. 20 kS/s

B. 25 kS/s

C. 30 kS/s

D. 32 kS/s

Correct Answer: D. 32 kS/s

Question 2

A 9 kHz sinusoid sampled at 10 kS/s appears at which baseband frequency?

A. 0.5 kHz

B. 1 kHz

C. 9 kHz

D. 19 kHz

Correct Answer: B. 1 kHz

Question 3

What is the principal purpose of an analog anti-aliasing filter?

A. To increase ADC resolution

B. To remove quantization noise after conversion

C. To attenuate unwanted analog components before sampling

D. To reconstruct a continuous waveform from samples

Correct Answer: C. To attenuate unwanted analog components before sampling

Question 4

For ideal uniform sampling, the sampled spectrum consists of copies of X(f) centered at:

A. Only f_s/2

B. Integer multiples of f_s

C. Only the signal carrier frequency

D. Random frequencies

Correct Answer: B. Integer multiples of f_s

Question 5

Which statement about aliasing is correct?

A. It can always be removed by a digital low-pass filter

B. It occurs only because of amplitude quantization

C. It is ambiguity caused by overlapping spectral replicas

D. It occurs only in nonuniform sampling

Correct Answer: C. It is ambiguity caused by overlapping spectral replicas

Question 6

The ideal zero-order-hold magnitude response contains a factor proportional to:

A. sinc(fT_s)

B. 1/f_s^2

C. sin(f_s t)

D. e^{fT_s}

Correct Answer: A. sinc(fT_s)

Question 7

A sinusoidal input has frequency 100 MHz and rms timing jitter is 1 ps. The approximate jitter-limited SNR is closest to:

A. 24 dB

B. 44 dB

C. 64 dB

D. 104 dB

Correct Answer: C. 64 dB

Question 8

Before decimation by M, a discrete-time signal must be:

A. High-pass filtered above the new Nyquist frequency

B. Low-pass filtered to prevent aliasing

C. Quantized with fewer bits

D. Converted to an analog signal

Correct Answer: B. Low-pass filtered to prevent aliasing

Question 9

Which condition is specifically associated with intentional bandpass sampling?

A. All frequencies must lie below f_s/2

B. Spectral replicas of the selected band must not overlap

C. No analog input filter is required

D. The sampling frequency must equal the carrier frequency

Correct Answer: B. Spectral replicas of the selected band must not overlap

Question 10

For an ideal N-bit uniform ADC with full-scale range V_FS, the approximate quantization step is:

A. V_FS × 2^N

B. V_FS/N

C. V_FS/2^N

D. 2^N/V_FS

Correct Answer: C. V_FS/2^N