Skip to content

DFT, FFT and Frequency Domain Analysis

Summary

This chapter focuses on frequency domain representation, spectral analysis, windowing techniques, and practical considerations for discrete-time frequency analysis.

Students will explore 10 key concepts that are essential for understanding this area of signal processing. This material provides the foundation for concepts introduced in later chapters.

Concepts Covered

This chapter covers the following 10 concepts from the learning graph:

  1. Frequency Domain
  2. Time Domain
  3. Spectrum
  4. Magnitude Spectrum
  5. Phase Spectrum
  6. Power Spectrum
  7. Spectral Analysis
  8. Spectral Leakage
  9. Window Functions
  10. Windowing Techniques

Prerequisites

This chapter builds on concepts from:

Introduction

Frequency domain analysis provides engineers and scientists with essential tools for understanding signal characteristics, designing filters, analyzing communication systems, and extracting information from complex data. While Chapter 6 introduced the mathematical foundations of Fourier transforms, this chapter focuses on practical application of discrete-time frequency analysis techniques including the DFT, FFT, and associated concepts of spectral representation. Moving from theory to practice requires understanding how finite-duration sampled signals map to frequency representations, how windowing affects spectral estimates, and how to interpret and manipulate frequency-domain data.

The distinction between time-domain and frequency-domain perspectives fundamentally shapes how engineers approach signal processing problems. Time-domain analysis reveals temporal structure, transient behavior, and instantaneous characteristics, while frequency-domain analysis exposes spectral content, filtering effects, and harmonic relationships. Many operations become simpler in one domain than the other, such as convolution (simple in frequency domain) versus frequency-selective filtering (conceptually simple in frequency domain, complicated in time domain). Mastering frequency domain techniques enables efficient problem solving and deep insight into signal behavior.

Domain Representations

Frequency Domain

The frequency domain represents signals through their spectral content, showing amplitude and phase as functions of frequency rather than time. This representation decomposes signals into constituent sinusoidal components, revealing which frequencies are present and their relative strengths. For a signal \(x(t)\) with Fourier transform \(X(f)\), the frequency-domain representation \(X(f)\) is generally complex-valued, encoding both magnitude and phase information at each frequency.

The frequency domain perspective proves particularly valuable for understanding filtering operations, where frequency-selective attenuation or amplification directly manifests as multiplication of the spectrum by the filter's frequency response. Communication systems engineers work extensively in the frequency domain to analyze bandwidth requirements, spectral efficiency, and interference characteristics. The frequency domain also facilitates analysis of periodic signals, where discrete spectral lines appear at harmonic frequencies rather than continuous spectra.

Time Domain

The time domain represents signals as functions of time, showing how signal amplitude varies from moment to moment. This familiar representation directly corresponds to physical measurements and waveform displays on oscilloscopes. For discrete-time signals, the time-domain representation \(x[n]\) consists of a sequence of samples indexed by integers, each representing the signal amplitude at a specific time instant.

Time-domain analysis excels at revealing transient behavior, rise and fall times, pulse characteristics, and temporal relationships between signals. Operations such as time shifting, time scaling, and amplitude modulation have intuitive time-domain interpretations. However, frequency-selective operations (filtering based on spectral content) often prove more complex in the time domain than in the frequency domain, motivating the use of transforms to shift between representations as convenient.

Spectral Representations

Spectrum

The spectrum of a signal describes its frequency content, typically showing the distribution of signal power or amplitude across frequency. For periodic signals, the spectrum consists of discrete components at the fundamental frequency and its harmonics (line spectrum), while aperiodic signals have continuous spectra. The spectrum provides essential information about signal characteristics including bandwidth, center frequency, and harmonic structure.

Spectral analysis enables engineers to identify frequency components present in complex signals, distinguish signals from noise based on spectral separation, and design filters that preserve desired frequency ranges while rejecting others. Spectrum analyzers are fundamental instruments in communications, audio engineering, vibration analysis, and numerous other fields where frequency content provides crucial diagnostic information.

Magnitude Spectrum

The magnitude spectrum \(|X(f)|\) or \(|X[k]|\) shows the amplitude of each frequency component without phase information, providing the most commonly visualized spectral representation. For complex-valued frequency-domain samples \(X[k] = A[k] + jB[k]\), the magnitude is computed as:

\[|X[k]| = \sqrt{A[k]^2 + B[k]^2}\]

The magnitude spectrum reveals which frequencies contain significant signal energy, enabling identification of dominant spectral components and estimation of signal bandwidth. Many applications care primarily about magnitude (which frequencies are present and how strong) rather than phase (temporal alignment of components), making magnitude spectra the default visualization for spectrum analyzers and spectral analysis software.

In audio applications, magnitude spectra show the tonal content of sounds, with peaks indicating fundamental frequencies and harmonics. In vibration analysis, magnitude spectra reveal resonant frequencies and mechanical characteristics. In communications, magnitude spectra indicate occupied bandwidth and spectral efficiency.

Phase Spectrum

The phase spectrum \(\angle X(f)\) or \(\angle X[k]\) shows the phase angle of each frequency component, computed as:

\[\angle X[k] = \arctan\left(\frac{B[k]}{A[k]}\right)\]

where appropriate quadrant determination ensures the correct angle. While often less visually striking than magnitude spectra, phase information proves crucial for applications requiring signal reconstruction, distinguishing minimum-phase from non-minimum-phase systems, and understanding group delay characteristics.

Phase linearity (constant slope of phase versus frequency) indicates constant group delay, preserving signal shape through filtering or transmission. Phase distortion manifests as non-constant group delay, causing different frequencies to experience different delays and potentially degrading signal quality. Telecommunications systems, audio processing for music production, and precision instrumentation all require careful attention to phase characteristics.

Power Spectrum

The power spectrum or power spectral density (PSD) shows signal power distribution versus frequency, particularly important for random signals where power provides a more stable measure than amplitude. For deterministic signals, the power spectrum is proportional to the squared magnitude spectrum, while for random processes, the PSD is defined through the Fourier transform of the autocorrelation function (Wiener-Khinchin theorem).

Power spectral density has units of power per Hertz, describing how signal power is distributed across the frequency spectrum. Integration of the PSD over a frequency band yields the total power in that band. In communications, PSD determines transmitter power requirements and receiver noise performance. In random vibration analysis, PSD characterizes excitation and response of mechanical systems to broadband inputs.

Spectral Analysis Techniques

Spectral Analysis

Spectral analysis encompasses methods for estimating and interpreting the frequency content of signals, ranging from simple FFT-based periodograms to sophisticated parametric and non-parametric estimators. The goal is to reveal spectral structure that enables signal identification, feature extraction, filtering, or understanding of underlying physical processes. Modern spectral analysis techniques address challenges including finite data length, noise contamination, time-varying spectra, and resolution limitations.

Non-parametric methods including the periodogram (squared magnitude of the FFT), Welch's method (averaged periodograms of overlapping segments), and multitaper methods provide direct spectral estimates without assuming particular signal models. Parametric methods including autoregressive (AR) modeling, moving average (MA) modeling, and ARMA modeling fit model parameters to data and derive spectra from the fitted models, often providing better resolution for short data records.

Spectral Leakage

Spectral leakage occurs when finite-duration observation windows cause signal energy at one frequency to spread into adjacent frequency bins, obscuring true spectral characteristics. This phenomenon arises because truncating a signal in time is equivalent to multiplying by a rectangular window, and the Fourier transform of a product equals convolution of the individual transforms. The rectangular window's transform (a sinc function with substantial sidelobes) convolves with the signal's true spectrum, spreading energy across frequencies.

Leakage manifests as broadened spectral peaks and elevated noise floors, making it difficult to distinguish closely-spaced frequency components or detect weak signals near strong ones. The severity of leakage depends on the window function used: rectangular windows produce the most leakage, while tapered windows (Hamming, Hanning, Blackman, etc.) reduce sidelobes at the cost of widening the main lobe and reducing frequency resolution.

Window Functions

Window functions taper time-domain signals to reduce spectral leakage by smoothly decreasing signal amplitude toward observation interval endpoints, minimizing discontinuities that cause leakage. Common window functions make different trade-offs between main lobe width (frequency resolution) and sidelobe level (leakage suppression). The choice of window depends on whether the application prioritizes resolving closely-spaced frequencies or detecting weak signals in the presence of strong ones.

Key window functions include:

  • Rectangular: No tapering, best frequency resolution, worst sidelobe suppression (~13 dB)
  • Hamming: \(w[n] = 0.54 - 0.46\cos(2\pi n/N)\), moderate characteristics, ~43 dB sidelobes
  • Hanning: \(w[n] = 0.5 - 0.5\cos(2\pi n/N)\), similar to Hamming, ~31 dB sidelobes
  • Blackman: Three-term cosine sum, excellent sidelobes (~58 dB), wider main lobe
  • Kaiser: Adjustable parameter controls sidelobe/resolution trade-off, very flexible

Window selection balances competing requirements based on signal characteristics and analysis objectives.

Windowing Techniques

Windowing techniques apply window functions to finite-duration signals before spectral analysis, implemented by pointwise multiplication of the signal by the window function. For a signal \(x[n]\) and window \(w[n]\), the windowed signal is \(x_w[n] = x[n]w[n]\). The DFT of the windowed signal provides the spectral estimate, with window characteristics determining the estimate's resolution and leakage properties.

Advanced windowing techniques include time-varying windows for spectrograms (short-time Fourier transform), multitapering (averaging spectra obtained with multiple orthogonal windows), and coherent versus non-coherent averaging. The proper application of windowing can dramatically improve spectral estimate quality, revealing spectral features that would be obscured by leakage with rectangular windows.

Windowing generally reduces the effective signal amplitude by a factor called the coherent gain or window gain, which must be compensated when accurate amplitude measurements are required. Similarly, the equivalent noise bandwidth of the window (wider than the frequency bin spacing) affects noise and power measurements.

Interactive Demonstrations

Diagram: Spectral Leakage and Windowing

MicroSim: Spectral Leakage and Windowing

Purpose

This interactive simulation demonstrates spectral leakage effects and how window functions mitigate leakage, showing the trade-offs between frequency resolution and sidelobe suppression.

Features

  • Generate test signals: single sinusoid, two close sinusoids, strong + weak sinusoid
  • Adjust signal frequency with fine control to observe leakage variation
  • Select window function: rectangular, Hamming, Hanning, Blackman, Kaiser
  • Display time-domain windowed signal
  • Show magnitude spectrum with leakage artifacts
  • Highlight main lobe width and sidelobe levels
  • Compare window characteristics: main lobe width, sidelobe level, scalloping loss
  • Adjustable FFT size to explore resolution effects

Learning Objectives

  • Understand spectral leakage as consequence of finite observation time
  • Recognize how window shape affects spectral estimates
  • Observe resolution versus leakage trade-offs
  • Learn when to use different window functions

Implementation Requirements

  • Canvas size: 670px wide, 625px total height (550px drawing + 75px controls)
  • Four plot panels: time signal, window function, windowed signal, magnitude spectrum
  • Signal frequency slider with fine adjustment
  • Window function dropdown selector
  • FFT size selector (256, 512, 1024, 2048 points)
  • Visual indicators for main lobe width and sidelobe level
  • Numerical readouts for window parameters
  • Cursor-based frequency and amplitude readout

Diagram: Time-Frequency Analysis Explorer

MicroSim: Time-Frequency Analysis Explorer

Purpose

This simulation demonstrates time-frequency analysis through spectrograms and short-time Fourier transforms, showing how signals with time-varying frequency content can be analyzed.

Features

  • Input signal options: chirp (frequency sweep), frequency-shift keying, music/speech sample
  • Adjustable STFT parameters: window length, overlap percentage, window type
  • Display spectrogram (color-coded time-frequency representation)
  • Show instantaneous spectrum at selected time instant
  • Display time-domain waveform with time cursor
  • Compare different window lengths to observe time-frequency resolution trade-off
  • Option to display magnitude or power spectrogram, linear or log scale

Learning Objectives

  • Understand time-frequency analysis for non-stationary signals
  • Observe uncertainty principle: time resolution versus frequency resolution
  • Recognize how STFT parameters affect analysis
  • Apply spectrograms to practical signals

Implementation Requirements

  • Canvas size: 670px wide, 625px total height (550px drawing + 75px controls)
  • Three plot areas: time waveform with cursor, spectrogram, instantaneous spectrum
  • Window length slider (affecting frequency vs. time resolution)
  • Overlap percentage slider (0-90%)
  • Window function selector
  • Color map selector for spectrogram
  • Scale options: linear/log for magnitude, linear/dB for amplitude
  • Cursor interaction for time selection and frequency readout

Comparison Tables

Domain Representation Mathematical Form Primary Applications
Time Amplitude vs. time \(x(t)\) or \(x[n]\) Waveform display, transient analysis, pulse measurements
Frequency Amplitude/phase vs. frequency \(X(f)\) or \(X[k]\) Spectral analysis, filter design, bandwidth analysis
Time-Frequency Amplitude vs. time and frequency \(S(t,f)\) (spectrogram) Non-stationary signal analysis, speech/music analysis
Spectrum Type Definition Units Typical Use
Magnitude Spectrum \(\|X[k]\|\) Signal units (V, Pa, etc.) Showing which frequencies are present
Power Spectrum \(\|X[k]\|^2\) Power units (\(V^2\), \(Pa^2\), etc.) Energy distribution across frequency
Power Spectral Density \(\lim_{\Delta f \to 0} \frac{P}{\Delta f}\) Power per Hz Random signal characterization, noise analysis
Phase Spectrum \(\angle X[k]\) Radians or degrees Group delay analysis, phase response
Window Function Main Lobe Width Sidelobe Level Scalloping Loss Best Use
Rectangular Narrowest (1.0 bins) Worst (-13 dB) 3.9 dB Maximum frequency resolution needed
Hamming Moderate (1.3 bins) Good (-43 dB) 1.8 dB General purpose, good all-around
Hanning Moderate (1.5 bins) Moderate (-31 dB) 1.4 dB General purpose, slightly less resolution
Blackman Wide (1.7 bins) Excellent (-58 dB) 1.1 dB Detecting weak signals near strong ones
Kaiser (\(\beta=5\)) Adjustable Adjustable (-50 dB typ) Variable Flexible, application-specific tuning
Flat Top Widest (3.8 bins) Good (-44 dB) 0.01 dB Accurate amplitude measurements
Analysis Method Frequency Resolution Time Resolution Computational Cost Best Application
FFT \(f_s/N\) \(N/f_s\) (full record) \(O(N \log N)\) Stationary signals, fast computation
STFT (Short-Time FFT) \(f_s/N_{win}\) Window duration \(O(N \log N_{win})\) per frame Time-varying signals, spectrograms
Welch's Method Depends on segment length N/A (averaging) Multiple FFTs Improved PSD estimates, smoothing
Multitaper Good (multiple windows) Full record Multiple FFTs High-quality PSD, low variance

Practical Considerations

When performing frequency-domain analysis on real signals, several practical issues warrant attention. Zero-padding (appending zeros to increase FFT length) provides finer frequency sampling without improving fundamental frequency resolution, which depends solely on observation duration. However, zero-padding can improve visualization of spectral features and interpolate between the natural frequency bins.

Frequency resolution equals the sampling rate divided by the number of samples (\(\Delta f = f_s/N\)), establishing the minimum frequency separation resolvable with a given data length. Improving frequency resolution requires longer observation windows, creating a fundamental trade-off with time resolution in time-varying signal analysis. The uncertainty principle quantifies this trade-off: \(\Delta t \cdot \Delta f \geq 1/(4\pi)\), showing that simultaneous high resolution in both domains is impossible.

Scaling and normalization conventions vary among FFT implementations, affecting amplitude interpretations. Some implementations include normalization factors in the forward transform, others in the inverse transform, and some split the factor equally. Understanding the particular convention used is essential for accurate amplitude and power measurements. Similarly, FFT outputs may be ordered differently (DC at index 0 versus centered), requiring appropriate interpretation or reordering.

Summary

This chapter has explored practical frequency-domain analysis techniques building on the theoretical Fourier foundations of Chapter 6. The distinction between time-domain and frequency-domain representations provides complementary perspectives on signal characteristics, with each domain offering advantages for particular operations and insights. Spectral representations including magnitude spectrum, phase spectrum, and power spectrum quantify different aspects of frequency content, enabling diverse applications from filter design to signal identification.

Spectral analysis faces challenges including spectral leakage caused by finite observation windows, addressed through judicious selection of window functions that trade off frequency resolution against sidelobe suppression. Common window functions including Hamming, Hanning, Blackman, and Kaiser provide different balances suitable for various applications, from resolving closely-spaced frequencies to detecting weak signals near strong ones. Understanding windowing effects and choosing appropriate parameters proves essential for high-quality spectral estimates.

Time-frequency analysis through techniques such as the short-time Fourier transform extends frequency analysis to non-stationary signals, revealing how spectral content evolves over time. The spectrogram visualization provides invaluable insights for speech processing, music analysis, radar signal processing, and any application involving signals whose frequency content changes. The uncertainty principle establishes fundamental limits on simultaneous time and frequency resolution, governing trade-offs in time-frequency analysis.

Mastering frequency-domain analysis techniques enables efficient signal processing, deep understanding of filter behavior, effective communication system design, and extraction of information from complex real-world signals across diverse engineering and scientific applications.