Computational Practice, Design Decisions, and Emerging Systems

Introduction

Signals and systems theory becomes engineering practice when mathematical models are converted into algorithms, hardware, software, and verified decisions. At postgraduate level, the central challenge is not merely to compute a Fourier transform or design a filter, but to decide which representation, approximation, architecture, sampling rate, numerical method, and validation strategy is appropriate for a stated purpose.

A practical signal-processing system is constrained by bandwidth, latency, memory, energy, cost, reliability, privacy, and the quality of available data. These constraints often conflict. Increasing sampling rate may simplify anti-aliasing but increase computation and storage. A longer filter may improve selectivity but increase delay. A neural model may improve classification accuracy but reduce interpretability and increase energy consumption. Engineering design therefore requires an explicit connection between system theory and system-level trade-offs.

This chapter develops a computational design framework for discrete-time and digital signal systems, examines implementation choices, and introduces emerging systems including adaptive, distributed, embedded, and learning-enabled signal-processing architectures.

Learning objectives

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

1. Translate a continuous-time signal problem into a computational model.
2. Select sampling, quantization, transform, and filtering strategies using stated assumptions.
3. Compare direct, transform-domain, fixed-point, floating-point, and hardware implementations.
4. Analyse computational complexity, latency, memory, numerical error, and robustness.
5. Design a validation workflow using analytical, simulation, and experimental evidence.
6. Explain the operation and limitations of adaptive, distributed, edge, and learning-enabled systems.
7. Make defensible design decisions under competing technical, economic, ethical, and regulatory constraints.

1. From a physical phenomenon to a computational system

A signal-processing design begins with a physical quantity such as acoustic pressure, acceleration, temperature, voltage, or biomedical activity. A useful abstraction is

Here, x_c(t) is a continuous-time signal, the sampler produces discrete-time samples, the quantizer maps amplitudes to finite numerical values, and the algorithm estimates, detects, reconstructs, or controls a process.

Every abstraction introduces assumptions. The sensor has a frequency response and noise; the sampler has a finite rate; the quantizer has finite resolution; the algorithm operates with finite word length; and the final decision may be affected by model mismatch. A rigorous design document should therefore state:

• the signal of interest and its units;
• the useful bandwidth and expected dynamic range;
• noise and interference conditions;
• allowable delay and failure probability;
• available computation, memory, power, and communication bandwidth;
• validation data and acceptance criteria.

A system specification should distinguish requirements from design choices. For example, “detect a fault within 50 ms with probability at least 0.95” is a requirement; “use a 64-tap FIR filter” is a design choice.

2. Sampling and quantization decisions

2.1 Sampling and aliasing

For a band-limited signal whose highest frequency component is f_max, ideal uniform sampling requires

where f_s is the sampling frequency in hertz. The strict inequality expresses the Nyquist condition for ideal reconstruction; in practical systems, margin is required because real filters are not brick-wall filters.

Sampling replicates the continuous-time spectrum at integer multiples of f_s. If spectral replicas overlap, aliasing occurs and cannot be removed by digital processing after sampling. Consequently, an analogue anti-aliasing filter is normally placed before the analogue-to-digital converter. If the signal contains energy up to 18 kHz and the transition band of the anti-aliasing filter is limited, selecting f_s=40 kHz rather than exactly 36 kHz provides practical guard band.

For a band-pass signal occupying frequencies from f_L to f_H, band-pass sampling may use a rate lower than 2f_H if spectral replicas do not overlap. The allowable rate depends on the band location, bandwidth, and a carefully verified alias pattern. It is not valid to apply the low-pass rule blindly or to assume that undersampling is automatically harmless.

2.2 Quantization and signal-to-quantization-noise ratio

An ideal uniform quantizer with B bits and full-scale peak-to-peak range V_pp has step size

where ∆ is measured in volts when V_pp is measured in volts. Under the assumptions of a uniformly distributed quantization error independent of the signal and no overload, the quantization-noise power is

For a full-scale sinusoid, the ideal signal-to-quantization-noise ratio is approximately

This result is a model, not a universal guarantee. Dither, converter nonlinearity, clipping, clock jitter, and non-sinusoidal signals alter the measured result. Increasing resolution is useful only if analogue noise and front-end imperfections do not already dominate.

Worked example: resolution and quantization noise

An 8-bit converter has a 2.56 V peak-to-peak input range. Then

∆=2.56/2⁸=0.010 V.

The ideal quantization-noise standard deviation is

σ_q=∆/√12≈2.89 mV.

The full-scale sinusoidal SQNR is approximately 6.02(8)+1.76=49.92 dB. If the input is only one-quarter of full-scale in amplitude, its power is one-sixteenth of the full-scale sinusoid, so the SQNR decreases by approximately 12.04 dB, assuming the same quantization-noise model and no other noise.

3. Computational representations and transforms

A signal can be represented in the time domain, frequency domain, transform domain, state space, or a learned feature space. Representation should be selected according to the operation required.

• Time-domain processing is natural for causal filtering, event detection, and time-local transients.
• Frequency-domain processing is useful for spectral analysis, convolution acceleration, and stationary-noise characterization.
• State-space models are effective for recursive estimation and control.
• Wavelet or multiresolution representations are useful when signal behaviour changes across time and scale.
• Learned representations can support classification but require data, validation, and governance.

For a length-N sequence, the discrete Fourier transform (DFT) is

The inverse DFT is

where x[n] is the input sequence, X[k] is its complex spectrum, j=√(−1), and N is the transform length. The DFT treats the finite record as one period of a periodic sequence. This assumption causes leakage when the record does not contain an integer number of sinusoidal cycles. A window function reduces leakage at the cost of wider main lobes and altered amplitude estimates.

The direct DFT requires order O(N²) arithmetic operations. The fast Fourier transform (FFT) exploits factorisation and symmetry to reduce the order to approximately O(N log₂ N) for suitable lengths. The FFT does not change the DFT result; it changes how efficiently that result is computed.

For a real signal, the spectrum has conjugate symmetry. One-sided spectra must be scaled carefully: except at DC and, for even N, the Nyquist bin, positive-frequency magnitudes generally require doubling when representing the total energy of the corresponding positive-negative pair.

4. Digital filter implementation and design decisions

A finite impulse response (FIR) filter is described by

where b_k are coefficients and the filter has M+1 taps. An FIR filter can be exactly linear phase when its coefficients have the required symmetry. It is inherently bounded-input bounded-output stable because its impulse response is finite.

An infinite impulse response (IIR) filter can be written as

with a₀=1 by convention. IIR filters often achieve a specified magnitude response with lower order than FIR filters, but their poles must lie inside the unit circle for causal BIBO stability. Quantization of coefficients can move poles and may cause instability, especially in high-order direct-form implementations.

Practical comparison

Second-order sections are commonly preferred for IIR implementation because they limit numerical sensitivity compared with one high-order direct-form polynomial. In real-time systems, coefficient scaling and internal overflow analysis are as important as the theoretical transfer function.

Convolution and FFT processing

For an FIR filter of length [[M]], direct convolution costs approximately order O(M) operations per output sample. For long filters, overlap-save or overlap-add FFT convolution can be more efficient. However, block processing introduces latency, memory overhead, and boundary-management requirements. The optimum block size is therefore a system decision rather than a purely mathematical one.

5. Numerical precision, complexity, and real-time feasibility

A theoretically correct algorithm can fail in implementation because of finite precision. Important error sources include:

• round-off and truncation;
• overflow and saturation;
• cancellation between nearly equal quantities;
• coefficient quantization;
• accumulation error in long sums;
• loss of significance in poorly conditioned problems.

Floating-point arithmetic provides a wide dynamic range but may consume more energy or hardware resources. Fixed-point arithmetic can be faster and more efficient, but requires explicit word-length, scaling, overflow, and quantization analysis. Saturation is often safer than wraparound because wraparound can convert a large positive value into a large negative value.

A basic real-time feasibility condition is

where T_proc includes acquisition, computation, memory transfer, scheduling, and communication time. For periodic frames, a stronger condition is that worst-case execution time, not merely average execution time, fits within the frame deadline. A design with low average latency but occasional missed deadlines may be unacceptable in control, medical, or safety-critical applications.

Useful engineering metrics include:

• operations per sample or per frame;
• memory footprint;
• throughput in samples per second;
• end-to-end latency;
• energy per sample or inference;
• numerical error relative to a reference implementation;
• probability of missed detection and false alarm.

Complexity analysis should identify dominant terms and constants. Two algorithms with the same asymptotic order can have very different performance because of memory locality, vectorisation, parallelism, and hardware acceleration.

6. A disciplined computational design workflow

A robust workflow is iterative:

1. Define the task. Specify estimation, detection, reconstruction, classification, or control objectives.
2. Characterise the signal. Measure bandwidth, dynamic range, nonstationarity, noise, and expected outliers.
3. Choose the sampling architecture. Set rate, anti-aliasing requirements, converter resolution, clock quality, and buffering.
4. Develop a reference model. Use high-precision or symbolic computation where possible.
5. Select an algorithm. Compare accuracy, complexity, latency, memory, and interpretability.
6. Map to implementation. Choose processor, FPGA, GPU, DSP, analogue front end, or hybrid architecture.
7. Verify progressively. Conduct unit, numerical, simulation, hardware-in-the-loop, and field tests.
8. Monitor after deployment. Record drift, failures, data distribution changes, and performance degradation.

The reference model must be independent enough to detect implementation errors. Comparing two implementations that share the same code defect is not meaningful validation. Test vectors should include nominal signals, boundary values, impulses, sinusoids, silence, overload, missing samples, timing irregularities, and adversarial or unexpected inputs.

7. Adaptive and state-estimation systems

Fixed filters assume that signal statistics and system characteristics are sufficiently stable. Adaptive systems update parameters from observed data. The least-mean-squares (LMS) adaptive filter illustrates the principle:

where w[n] is the coefficient vector, x[n] is the input vector, e[n] is the error, and µ is the step size. For the standard LMS algorithm, a commonly used sufficient practical condition is

where λ_max is the largest eigenvalue of the input correlation matrix. In practice, a smaller step size improves stability and reduces misadjustment but slows convergence. The bound depends on the signal statistics and implementation assumptions.

The normalized LMS algorithm scales the update by input energy:

where ε>0 prevents division by zero. Adaptive filtering is used in echo cancellation, active noise control, equalisation, channel tracking, and biomedical interference removal.

State-space estimation provides another adaptive framework. A linear discrete model is

where x_k is the hidden state, u_k is the input, y_k is the measurement, and w_k, v_k are process and measurement noise. The Kalman filter is optimal in the minimum mean-square-error sense for linear models with Gaussian noise and known statistics. Outside these assumptions it remains useful, but its optimality claim no longer automatically applies.

8. Distributed, edge, and networked signal systems

Modern sensing systems frequently distribute computation across sensors, edge devices, gateways, and cloud services. A centralised architecture sends raw data to one processor; an edge architecture performs feature extraction or decisions near the sensor; a distributed architecture allows multiple nodes to estimate a common quantity or cooperate on inference.

Edge processing reduces communication load and may reduce latency and privacy exposure. It also imposes limitations on memory, energy, model size, and update mechanisms. Distributed systems introduce clock synchronisation, packet loss, quantisation, communication delay, and consensus problems. A locally accurate estimate may be globally inconsistent if nodes use different calibration or stale information.

A design should specify whether the system requires:

• synchronised or asynchronous sampling;
• raw-data retention or feature-only transmission;
• deterministic or best-effort latency;
• centralised, federated, or decentralised learning;
• graceful degradation when nodes fail.

Federated learning can keep raw data on devices while sharing model updates, but it does not guarantee privacy by itself. Updates may leak information, and model poisoning remains possible. Secure aggregation, differential privacy, authentication, and access control may be required according to the threat model.

9. Learning-enabled signal processing

Machine learning is valuable when the mapping from signal patterns to decisions is difficult to express analytically or when large labelled datasets are available. Typical tasks include modulation recognition, speech enhancement, fault diagnosis, image reconstruction, and anomaly detection.

A responsible design compares a learned model with a strong signal-processing baseline. A neural classifier may achieve high test accuracy but fail under sensor drift, changed noise, class imbalance, or distribution shift. Performance should therefore be reported using task-appropriate measures such as precision, recall, receiver operating characteristic curves, calibration, detection delay, and uncertainty—not accuracy alone.

Training, validation, and test data must be separated by time, subject, device, or environment when random splitting would cause leakage. For a rare-event detector, accuracy can be misleading. If only 1% of records contain faults, a classifier that always predicts “no fault” achieves 99% accuracy but has zero recall for faults.

Hybrid systems often provide a better engineering compromise: conventional filtering and physical constraints reduce noise and enforce known structure, while machine learning handles residual classification or nonlinear estimation. Explainability is especially important where decisions affect safety, health, employment, or access to services.

10. Case study: edge-based bearing-fault monitoring

The following is an illustrative engineering case, not a report of a particular commercial installation.

Background

A manufacturing plant wants to detect early bearing faults in induction-motor drives. Accelerometers mounted near the bearings produce vibration signals. The system must raise an alarm within 2 s, operate for one year on limited maintenance power, and continue producing a safe health indicator when network connectivity is lost.

Problem analysis

The useful vibration content extends to approximately 12 kHz, but electromagnetic interference and structural resonances create strong out-of-band components. The first design decision is sampling. A rate of 25.6 kHz provides a nominal Nyquist frequency of 12.8 kHz, leaving only a narrow transition band. The design therefore uses an analogue anti-aliasing filter and selects 32 kHz to provide practical margin. A 16-bit converter is selected after measuring that analogue noise, rather than quantisation, is expected to dominate below the sensor’s full-scale range.

The raw data rate per sensor is

32,000×16=512,000 bits/s,

or 64 kB/s before framing and protocol overhead. Continuous transmission from many sensors is undesirable, so computation is performed at the edge.

Approach

Each 1,024-sample frame represents

1024/32,000=0.032 s,

or 32 ms. Frames are windowed, transformed using an FFT, and converted into spectral-band energy features. A calibrated baseline is used to compensate for sensor sensitivity. A compact classifier combines these features with a trend model. The device stores recent features and alarm evidence, while raw waveforms are retained only around confirmed events.

The 2-s decision requirement permits approximately 62 frames, but the implementation uses a shorter aggregation interval to allow confirmation and communication margin. A watchdog detects processor lockup; a missing-data flag prevents the classifier from interpreting communication failure as a healthy machine.

Validation

Testing includes healthy bearings, seeded defects in a laboratory rig, variable motor load, temperature changes, sensor replacement, and network interruption. Metrics include false alarms per operating hour, detection probability at a specified defect severity, detection delay, energy per frame, and recovery time after reset. A held-out machine—not merely a held-out random frame—is used for final evaluation to test generalisation across assets.

Outcome and learning points

The design illustrates several principles. Sampling rate is chosen with the anti-aliasing transition band in mind; edge processing is justified by bandwidth and availability requirements; a learned classifier is supported by calibrated signal features and explicit missing-data handling; and validation reflects deployment conditions rather than only laboratory accuracy. The most important outcome is not a particular algorithm but a traceable chain from requirement to measurement, implementation, and evidence.

11. Design governance, ethics, and reliability

Signals may contain personal, medical, industrial, or location information. Data minimisation, purpose limitation, retention controls, access management, and secure transmission should be incorporated at the architecture stage. Consent and legal authority depend on jurisdiction and application.

Safety-critical systems require hazard analysis, independent verification, traceability, fail-safe behaviour, and controlled change management. A signal-processing model should not silently change after deployment without versioning and revalidation. AI-enabled systems additionally require documentation of training data, known limitations, performance across relevant populations, and mechanisms for human review.

Environmental cost is also a design concern. Sampling, communication, and inference consume energy. A lower-complexity model, duty-cycled sensor, or event-triggered architecture may provide a better lifecycle outcome than a permanently active high-performance model. Security threats include spoofed sensor signals, replay attacks, malicious firmware, data poisoning, and denial of service. Threat modelling should be performed before architecture is frozen.

12. Contemporary directions and research challenges

Several developments are reshaping computational signals and systems:

• Neuromorphic and event-based sensing: sensors transmit changes rather than uniform-rate samples, reducing data for sparse events but complicating reconstruction and algorithms.
• In-memory and near-memory computing: reduces data movement, which can dominate energy in large models.
• Differentiable signal-processing systems: jointly learn selected filter or transform parameters while retaining interpretable structure.
• Physics-informed learning: incorporates conservation laws, known dynamics, or measurement models into estimation.
• Digital twins: combine sensor streams with dynamic system models for prediction and maintenance.
• 6G and integrated sensing and communication: communication waveforms are designed to support both data transfer and environmental inference.
• Tiny machine learning: compresses models through pruning, quantisation, knowledge distillation, and architecture search for microcontrollers.

Open research questions include reliable uncertainty quantification under distribution shift, joint optimisation of sensing and computation, trustworthy federated learning, robust operation under adversarial interference, and formal verification of learning-enabled components. These areas do not eliminate classical signals-and-systems theory; they make modelling, sampling, stability, estimation, and validation even more important.

Best-practice checklist

A technically mature computational design should:

1. State signal, system, and performance assumptions explicitly.
2. Prevent aliasing before conversion and document converter limits.
3. Distinguish theoretical performance from measured performance.
4. Analyse worst-case latency, memory, overflow, and failure modes.
5. Use an independent reference implementation and reproducible test data.
6. Select FIR, IIR, FFT, adaptive, or learned methods according to requirements.
7. Validate across devices, environments, operating conditions, and time.
8. Monitor deployed systems for drift and degradation.
9. Protect sensitive data and secure update and communication paths.
10. Preserve human oversight where errors can cause significant harm.

Chapter Summary

Computational practice in signals and systems is the disciplined conversion of physical measurements into reliable numerical decisions. Sampling and quantization establish the information available to computation; anti-aliasing, resolution, and clock quality must be considered together. DFT and FFT representations support analysis and efficient convolution, while FIR and IIR filters offer different compromises among phase, order, stability, and numerical sensitivity.

Implementation requires analysis of finite precision, complexity, memory, latency, energy, and worst-case execution time. Adaptive filters and state estimators address changing environments, while distributed and edge architectures trade communication against local computation and privacy. Learning-enabled systems can extend conventional methods but require careful data separation, uncertainty analysis, robustness testing, and governance.

The central professional skill is defensible design decision-making: connecting requirements to assumptions, algorithms, hardware, verification evidence, and responsible deployment. Emerging systems—including event-based sensors, physics-informed learning, digital twins, and integrated sensing and communication—build on rather than replace the foundational concepts of signals and systems.

Key Terms

Sampling: Conversion of a continuous-time signal into values at discrete time instants.

Aliasing: Irreversible spectral overlap caused by insufficient sampling or inadequate anti-aliasing.

Quantization: Mapping of continuous-amplitude samples to a finite set of numerical levels.

Quantization noise: Error component introduced by ideal amplitude quantization.

DFT: Finite-length discrete Fourier transform of a sequence.

FFT: Efficient algorithm for computing the DFT.

Spectral leakage: Spreading of finite-record spectral energy across frequency bins.

FIR filter: Filter with a finite-duration impulse response.

IIR filter: Recursive filter with, in general, an infinite-duration impulse response.

Fixed-point arithmetic: Numerical representation with a fixed scale and finite word length.

Floating-point arithmetic: Numerical representation using significand and exponent fields.

Latency: Time between an input event and the corresponding output or decision.

Adaptive filter: Filter whose parameters update using observed data.

LMS algorithm: Gradient-based adaptive filtering method using an instantaneous error.

Kalman filter: Recursive state estimator for dynamic systems under specified statistical assumptions.

Edge computing: Processing performed near the data source rather than exclusively in a remote cloud.

Distribution shift: Change between the data conditions used for development and deployment.

Federated learning: Collaborative model training in which raw data remains distributed.

Digital twin: Computational representation linked to a physical system through data and models.

Hardware-in-the-loop: Testing in which real hardware interacts with a simulated environment.

Review Questions

Short Answer Questions

1. Why must anti-aliasing generally be performed before analogue-to-digital conversion?
2. State two differences between FIR and IIR filters.
3. What assumptions underlie the ideal quantization-noise model?
4. Why can average execution time be insufficient for a real-time design?
5. What is the purpose of an independent reference implementation?

Long Answer Questions

1. Explain how sampling rate, anti-aliasing transition band, converter resolution, and signal bandwidth should be selected as a unified design problem.
2. Compare direct convolution and FFT-based convolution, including computational, memory, and latency trade-offs.
3. Discuss finite-precision effects in IIR filters and explain why second-order sections are often preferred.
4. Explain the role of adaptive filtering and Kalman estimation, including their assumptions and limitations.
5. Evaluate the benefits and risks of deploying a learning-enabled signal-processing system at the edge.

Analytical / Application Questions

1. A vibration signal contains useful components up to 10 kHz, while an anti-aliasing filter requires a 2 kHz transition band. Propose a sampling rate and justify the guard band and filter requirements.
2. Design a validation plan for a medical signal detector whose false-alarm rate must remain below a specified threshold across multiple hospitals. Identify data splits, metrics, stress tests, and governance controls.
3. For an industrial monitoring system with limited battery energy and intermittent connectivity, compare centralised cloud processing, local edge processing, and a hybrid architecture. Recommend one and defend the decision using bandwidth, latency, reliability, privacy, and maintenance criteria.

Multiple Choice Questions with Answers

Question 1

For an ideal low-pass signal with highest frequency f_max, which condition is required for ideal uniform sampling?

A. f_s < f_max

B. f_s = f_max

C. f_s > 2f_max

D. f_s > f_max/2

Correct Answer: C. f_s > 2f_max

Question 2

An ideal 8-bit quantizer has a 2.56 V peak-to-peak range. What is its quantization step size?

A. 1.00 mV

B. 10.0 mV

C. 80.0 mV

D. 256 mV

Correct Answer: B. 10.0 mV

Question 3

Which statement about the FFT is correct?

A. It changes the DFT definition

B. It computes the DFT more efficiently

C. It eliminates spectral leakage

D. It is applicable only to analogue signals

Correct Answer: B. It computes the DFT more efficiently

Question 4

Which property is inherently associated with a finite-length FIR filter?

A. BIBO stability

B. Unstable poles

C. Infinite memory

D. Mandatory nonlinear phase

Correct Answer: A. BIBO stability

Question 5

For a causal IIR filter, which pole condition is generally required for BIBO stability?

A. All poles outside the unit circle

B. All poles on the imaginary axis

C. All poles inside the unit circle

D. At least one pole at infinity

Correct Answer: C. All poles inside the unit circle

Question 6

In the LMS update, reducing the step size generally produces which trade-off?

A. Faster convergence and greater misadjustment

B. Slower convergence and potentially lower misadjustment

C. No effect on convergence

D. Guaranteed instability

Correct Answer: B. Slower convergence and potentially lower misadjustment

Question 7

Why is edge processing often used in distributed sensing systems?

A. It always increases raw-data transmission

B. It can reduce communication load and latency

C. It eliminates the need for sensors

D. It guarantees privacy without security controls

Correct Answer: B. It can reduce communication load and latency

Question 8

A classifier applied to data with severe class imbalance should not be evaluated using accuracy alone because:

A. Accuracy cannot be computed for classifiers

B. A majority-class predictor may appear highly accurate while missing rare events

C. Accuracy is always lower than recall

D. Class imbalance affects only analogue circuits

Correct Answer: B. A majority-class predictor may appear highly accurate while missing rare events

Question 9

What is a principal reason for using second-order sections in high-order IIR implementation?

A. They remove all computation

B. They reduce numerical sensitivity compared with one high-order direct form

C. They guarantee linear phase

D. They require no coefficient quantization

Correct Answer: B. They reduce numerical sensitivity compared with one high-order direct form

Question 10

Which statement best describes federated learning?

A. All raw data must be uploaded to one server

B. Training can occur across devices while raw data remains distributed

C. It is identical to analogue filtering

D. It guarantees immunity to model poisoning

Correct Answer: B. Training can occur across devices while raw data remains distributed