Frequency and Time Domains

The Central Idea

Suppose a signal is represented by a function \(x(t)\). In the time domain, we study \(x(t)\) directly. Under a Fourier transform convention,

the same signal is represented in the frequency domain by \(X(\omega)\). The inverse transform reconstructs the original signal:

Thus,

The transform changes the coordinates used to describe the object. It does not create a different signal.

Two Views of the Same Information

The time domain asks questions such as:

• When does the signal begin?
• How rapidly does it rise or fall?
• Where do pulses, transients, or discontinuities occur?
• What is the value at a particular instant?
• How does the output evolve after an input is applied?

The frequency domain asks questions such as:

• Which oscillatory frequencies are present?
• Which frequencies carry the most amplitude or energy?
• What bandwidth does the signal occupy?
• Which frequencies are amplified, attenuated, or phase shifted by a system?
• Does a periodic pattern contain a fundamental frequency and harmonics?

Choosing a domain is therefore a representation decision. We choose the view that makes the current structure easiest to analyze.

What the Time Domain Emphasizes

A time-domain graph places time on the horizontal axis and signal value on the vertical axis. It directly displays events and evolution.

For example, a decaying exponential

shows an immediate start at \(t=0\) followed by exponential decay. This description makes causality and decay rate visually natural.

The time domain is particularly effective for pulses, switching events, transient behavior, delays, peaks, settling, and localized disturbances.

What the Frequency Domain Emphasizes

The frequency domain reorganizes the same signal according to oscillatory content. Instead of asking what happens at a particular instant, we ask how much of each frequency is needed to reconstruct the signal.

For a transform \(X(\omega)\), two central quantities are

and

The magnitude spectrum \(|X(\omega)|\) describes the strength of spectral components, while the phase spectrum \(\arg X(\omega)\) describes their relative alignment.

A Single Sinusoid

Consider

In the time domain, this is an oscillating waveform with amplitude \(A\), angular frequency \(\omega_0\), and phase \(\phi\).

Euler's identity gives

Spectrally, the sinusoid is concentrated at the two angular frequencies \(\omega=\pm\omega_0\). The time-domain oscillation and the frequency-domain spectral lines are two descriptions of the same structure.

Angular Frequency and Ordinary Frequency

Frequency may be measured in cycles per second, or hertz, using \(f\), or in radians per second using angular frequency \(\omega\). They are related by

A frequency of \(f=10\text{ Hz}\) corresponds to

Transform formulas must be read with attention to which convention is being used. A missing factor of \(2\pi\) is often a convention mismatch rather than a deeper mathematical disagreement.

Periodic Signals and Discrete Frequencies

A periodic signal with period \(T\) has fundamental angular frequency

A Fourier series represents the signal using integer multiples of this fundamental:

Thus a periodic signal typically produces a discrete line spectrum. The lines may include a constant component, a fundamental, and higher harmonics.

Aperiodic Signals and Continuous Frequencies

A nonperiodic signal on an unbounded time axis generally requires a continuous range of frequencies. The Fourier transform replaces the discrete Fourier-series coefficients with the continuous spectral function \(X(\omega)\).

This distinction gives a useful organizational rule:

Sampling and periodic extension introduce additional details, but this pattern is a strong first guide.

Amplitude and Phase Both Matter

A complex spectrum can be written in polar form as

The magnitude \(|X(\omega)|\) and phase \(\phi(\omega)\) encode different information. Two signals can have the same magnitude spectrum but different phases and therefore look very different in time.

It is therefore generally incorrect to treat the magnitude spectrum alone as a complete description of the signal.

Real Signals and Conjugate Symmetry

If \(x(t)\) is real-valued, its Fourier transform satisfies

Consequently,

so the magnitude spectrum is even. The phase has the corresponding odd symmetry wherever phase is defined consistently.

This symmetry explains why positive-frequency information often determines the negative-frequency portion for real signals.

Example: A Rectangular Pulse

Let

Then

Evaluating gives

Equivalently,

when \(\operatorname{sinc}(z)=\sin z/z\). A signal that is sharply bounded in time therefore spreads across many frequencies.

Duration and Bandwidth

The rectangular-pulse example illustrates a broad reciprocal relationship. Making a pulse narrower in time generally spreads its spectrum. Making it wider in time generally concentrates its spectrum.

Fourier scaling makes this precise. If

then for nonzero \(a\),

Compressing the signal in time expands the spectrum in frequency. Stretching it in time compresses the spectrum.

Time-Frequency Localization

A signal cannot generally be made arbitrarily concentrated in both time and frequency. A very localized event requires a broad collection of frequencies, while a nearly pure frequency must persist over a long time.

With suitable definitions of time spread \(\sigma_t\) and angular-frequency spread \(\sigma_\omega\), a Fourier uncertainty relation takes the form

The exact numerical constant depends on the transform convention and definitions used, but the conceptual message is stable: localization in one domain requires spread in the other.

Time Shifts Become Phase Factors

If

then delaying the signal by \(t_0\) gives

A pure time delay leaves the magnitude spectrum unchanged:

The delay appears entirely as a frequency-dependent phase shift.

Frequency Shifts Become Modulation

Multiplying a time signal by a complex exponential shifts its spectrum. If

then

This property is called modulation or frequency translation. It is fundamental in communication systems because information concentrated near one frequency can be shifted to another spectral band.

Real Modulation Produces Two Shifted Copies

Since

multiplication by a cosine gives

A baseband spectrum is copied and shifted to positive and negative carrier frequencies.

Differentiation in Time Becomes Multiplication in Frequency

Under suitable conditions,

Each derivative introduces another factor of \(i\omega\):

This is one reason transform methods simplify differential equations. A differentiation operation in time becomes algebraic multiplication in frequency.

Why Differentiation Emphasizes High Frequencies

The factor \(|\omega|\) grows with frequency. Therefore, differentiation tends to amplify high-frequency spectral components relative to low-frequency ones.

This provides a spectral explanation for a familiar numerical fact: derivatives can be sensitive to rapid fluctuations and measurement noise. High-frequency noise may be small in the original signal but much more prominent after differentiation.

Integration Suppresses High Frequencies

Formally, integration corresponds to division by \(i\omega\), together with care at \(\omega=0\) and with constants of integration or distributional terms. Away from zero frequency, the basic relationship is

Because \(1/|\omega|\) decreases as \(|\omega|\) grows, integration tends to attenuate rapid oscillations. This is one reason accumulation often appears smoother than differentiation.

Convolution in Time Becomes Multiplication in Frequency

If

then the Fourier convolution theorem gives

This is one of the most important bridges between the two domains. A potentially complicated integral operation in time becomes ordinary pointwise multiplication in frequency.

Multiplication in Time Becomes Convolution in Frequency

The dual relationship is also important. Under the convention used here,

Thus multiplication that looks simple in time can spread or mix spectral content. Windowing is a major example: multiplying a signal by a finite-time window convolves its ideal spectrum with the transform of that window.

System Response in the Two Domains

For a linear time-invariant system with impulse response \(h(t)\), the output is

In the frequency domain,

The function \(H(\omega)\) is the system's frequency response. It tells us how each complex-exponential frequency component is scaled and phase shifted.

Complex Exponentials Are Special Inputs

Let the input to an LTI system be

Then

Factor out the part independent of \(\tau\):

The integral is \(H(\omega_0)\), so

The input keeps its form. Only its complex amplitude changes. This is precisely an eigenfunction relationship and will become a major bridge to the next lesson.

Magnitude Response and Phase Response

Write the frequency response as

Then \(|H(\omega)|\) controls gain or attenuation at each frequency, while \(\theta(\omega)\) controls phase shift.

For a component at frequency \(\omega_0\), the system multiplies its magnitude by \(|H(\omega_0)|\) and adds \(\theta(\omega_0)\) to its phase.

Low-Pass Filtering

A low-pass filter preserves low-frequency components more strongly than high-frequency components. Ideally,

In the time domain, such filtering often appears as smoothing or averaging. In the frequency domain, it appears as suppression of the high-frequency portion of the spectrum.

High-Pass Filtering

A high-pass filter does the opposite: it suppresses slowly varying or constant components and preserves more rapid variation.

This can emphasize edges, transitions, and sudden changes. The time-domain behavior may resemble a differencing or differentiation operation, consistent with the factor \(i\omega\) associated with differentiation.

Band-Pass and Band-Stop Filtering

A band-pass filter preserves a selected interval of frequencies and attenuates frequencies outside that interval. A band-stop or notch filter attenuates a selected interval while preserving frequencies outside it.

These descriptions are especially natural in the frequency domain because the desired action can be stated directly in terms of spectral location.

Example: Removing a High-Frequency Component

Suppose

The time-domain signal contains a slow oscillation plus a much faster one. Its spectrum has components at \(\omega=\pm2\) and \(\omega=\pm40\).

A low-pass filter that passes frequencies near \(|\omega|=2\) and strongly attenuates frequencies near \(|\omega|=40\) produces approximately

The filtering decision is easier to describe spectrally than by editing the waveform point by point in time.

Bandwidth

The bandwidth of a signal or system describes a range of frequencies that meaningfully contributes to the signal or is effectively transmitted by the system.

For an idealized band-limited signal,

The number \(\Omega_B\) is then a natural angular-frequency bandwidth limit. Practical spectra rarely vanish exactly outside a perfect interval, so real applications use operational definitions such as cutoff levels or energy percentages.

DC and the Zero-Frequency Component

The frequency \(\omega=0\) corresponds to a constant component. In periodic Fourier analysis, it is often called the DC component.

For a periodic signal, the zero-frequency coefficient is related to its average value. A nonzero average therefore appears as spectral content at zero frequency.

Energy Is Preserved Across Representations

Parseval's or Plancherel's relation expresses the fact that transform-domain representation does not destroy signal energy. Under the convention used here,

The scale factor reflects the Fourier-transform convention. Conceptually, the same energy can be measured in the time representation or in the frequency representation.

Energy Spectral Density

For a finite-energy signal, \(|X(\omega)|^2\) describes how spectral energy is distributed across frequency, up to the normalization used by the transform.

Large values of \(|X(\omega)|^2\) identify frequency regions that contribute strongly to the signal's total energy.

Time-Domain Delay versus Frequency-Domain Phase

Suppose an output is simply a delayed version of an input:

In frequency,

The magnitude ratio is

while the phase is linear:

A pure delay is therefore easy to identify as linear phase in the frequency domain.

Phase Distortion

If a system changes different frequencies by different phase relationships, a waveform can be reshaped even when the magnitude response is nearly flat.

Preserving amplitude alone does not guarantee preservation of waveform shape. Relative phase determines how frequency components line up when reconstructed in time.

Sampling Connects Continuous and Discrete Domains

A continuous-time signal \(x(t)\) can be sampled at times

producing the discrete sequence

The sampling frequency is

Sampling creates a bridge between continuous Fourier analysis and discrete spectral computation, but it also introduces periodic spectral replication and the possibility of aliasing.

The Nyquist Condition

For an ideal signal band-limited to ordinary frequencies \(|f|\le B\), perfect reconstruction requires, under the usual sampling theorem assumptions,

The quantity \(f_s/2\) is the Nyquist frequency. Frequencies above this limit can fold into lower apparent frequencies after sampling.

Aliasing Is a Frequency-Domain Collision

After sampling, frequency components separated by integer multiples of the sampling frequency can become indistinguishable in the discrete data.

This is why aliasing is especially transparent in the frequency domain: replicated spectra overlap. In the time-domain samples, the resulting error may look like a completely different oscillation.

Finite Observation Changes the Spectrum

In practice, signals are observed for only a finite duration. Recording a signal over a finite interval is equivalent to multiplying an ideal infinite-duration signal by a window function.

Since multiplication in time becomes convolution in frequency, the observed spectrum is not simply the ideal spectrum cut off at the endpoints. It is convolved with the spectrum of the window.

Spectral Leakage

Spectral leakage occurs when energy from an ideal frequency component spreads into neighboring frequency bins in a finite-data Fourier analysis.

Leakage is not usually evidence that the original signal physically contains all of those nearby frequencies. It is often a consequence of finite observation and the chosen window.

Windowing Is a Tradeoff

A window can reduce abrupt endpoint discontinuities and suppress spectral sidelobes, but this typically broadens the main spectral lobe. Window selection therefore trades one kind of frequency-domain resolution against another.

This is another example of the broader rule that operations that improve localization in one representation can reduce localization in the other.

Frequency Resolution for Finite Data

If \(N\) samples are collected at sampling rate \(f_s\), the observation duration is approximately

The DFT frequency-bin spacing is

Longer observation time produces finer frequency spacing. This is a practical manifestation of the time-frequency reciprocity.

Zero Padding Does Not Create New Physical Resolution

Appending zeros before computing a longer DFT samples the same underlying finite-record spectrum at more frequency points. It can make plots smoother and improve numerical interpolation of spectral peaks.

It does not provide the same information as collecting the signal for a longer time. True ability to separate nearby frequencies depends primarily on the observation interval and window behavior.

When the Time Domain Is Usually Better

The time domain is often the more efficient representation when the problem is centered on:

• arrival time and delay;
• transient response;
• switching and impulses;
• peak amplitude;
• settling behavior;
• causal sequence of events;
• local waveform shape;
• direct numerical simulation in time.

When the Frequency Domain Is Usually Better

The frequency domain is often the more efficient representation when the problem is centered on:

• periodicity and harmonics;
• resonance;
• bandwidth;
• filter design;
• steady-state sinusoidal response;
• noise concentrated in particular spectral regions;
• convolution and LTI systems;
• differential operators that become algebraic multipliers.

Neither Domain Is Inherently More Real

Physical measurements are often collected as functions of time or position, so the time or spatial domain may feel more direct. Frequency components, however, are not merely computational inventions. They correspond to measurable oscillatory behavior and to the natural responses of many physical systems.

The two domains should be treated as complementary coordinate systems for the same mathematical object.

A Domain-Translation Table

Several major Fourier relationships can be summarized as follows.

A large part of transform problem solving consists of recognizing which side of this table makes the desired operation simpler.

Example: Solving a Filtering Problem by Changing Domains

Suppose an LTI system has impulse response \(h(t)\) and receives input \(x(t)\). Directly in time,

If the Fourier transforms are easy to obtain, we may instead compute

multiply,

and then invert:

The answer is the same. Only the computational route has changed.

Example: Reading a Spectrum

Suppose a measured signal has strong spectral peaks at

One possible interpretation is a fundamental frequency of \(50\text{ Hz}\) with second and third harmonics. The frequency domain makes this harmonic relationship immediate because

In a complicated time-domain waveform, the same structure may be much harder to recognize visually.

Example: A Gaussian Is Localized in Both Domains

Consider a Gaussian

Its Fourier transform is another Gaussian:

Increasing \(a\) makes the time-domain Gaussian narrower while making the frequency-domain Gaussian wider. Decreasing \(a\) has the opposite effect.

Gaussians are especially important because they achieve the minimum spread permitted by the standard Fourier uncertainty relation.

From Frequency Analysis to Spectral Methods

The frequency-domain viewpoint is one instance of a more general strategy: represent an object in a basis whose elements interact simply with the operator or system under study.

For LTI systems, complex exponentials are particularly effective because

is reproduced in the same form after the system acts, multiplied only by \(H(\omega)\). This is exactly the behavior of an eigenfunction.

Frequency Is One Kind of Spectral Variable

The word spectrum is broader than ordinary physical frequency. In linear algebra, eigenvalues form a spectrum. In differential equations, boundary conditions can select discrete eigenvalues and corresponding eigenfunctions. In quantum mechanics, spectral values can represent observable quantities such as energy.

Fourier analysis is therefore part of a larger spectral viewpoint: decompose an object into modes adapted to the operator or geometry of the problem.

Common Error: Thinking the Transform Changes the Signal

A transform changes representation, not the underlying mathematical information. If the transform is invertible, then

means that either representation can be recovered from the other under the appropriate conditions.

Common Error: Ignoring Phase

Plotting only \(|X(\omega)|\) can be useful, but magnitude alone is generally not enough to reconstruct \(x(t)\). Phase controls how spectral components align in time.

A signal delay, for example, can leave the magnitude spectrum exactly unchanged while altering the phase spectrum.

Common Error: Confusing Hz and rad/s

Ordinary frequency \(f\) and angular frequency \(\omega\) satisfy

A cutoff quoted as \(100\text{ Hz}\) is not numerically the same as a cutoff quoted as \(100\text{ rad/s}\).

Common Error: Treating Every Spectral Peak as a Separate Physical Source

Harmonics, modulation sidebands, finite-window leakage, resonances, and aliasing can all produce spectral structure. A peak in a computed spectrum must therefore be interpreted in context.

Spectral analysis reveals structure, but physical interpretation still requires a model of how the data were generated and measured.

Common Error: Believing More DFT Points Always Mean More Information

Zero padding can produce a denser frequency grid, but it does not replace a longer observation interval. It interpolates the transform of the available record rather than adding new measured behavior.

Common Error: Forgetting the Domain When Multiplying

The equation

corresponds to convolution in time, not ordinary multiplication in time. Conversely, ordinary multiplication in time corresponds to convolution in frequency, with the normalization dictated by the transform convention.

A Reliable Procedure for Moving Between Domains

1. Identify the operation or structure that makes the problem difficult.
2. Ask whether that operation becomes simpler under a transform.
3. State the transform convention being used.
4. Transform the known signals, functions, or equations.
5. Perform the simpler algebraic or spectral operation.
6. Track magnitude, phase, units, and normalization carefully.
7. Apply the inverse transform when a time-domain answer is required.
8. Check the result using symmetry, limiting behavior, energy, dimensions, or direct substitution.

Practice 1: Convert Frequency Units

A sinusoid has ordinary frequency \(25\text{ Hz}\). Find its angular frequency.

Practice 2: Identify the Fundamental Frequency

A periodic signal has period

Find its fundamental frequency in hertz.

Practice 3: Time Shift

If

find the Fourier transform of \(x(t-3)\).

Practice 4: Frequency Shift

If

find the transform of

Practice 5: Differentiation

If \(x(t)\longleftrightarrow X(\omega)\), find the Fourier transform of

Practice 6: LTI Output

A system has frequency response

If the input spectrum is \(X(\omega)\), write the output spectrum.

Practice 7: Frequency Resolution

A data record contains \(N=2000\) samples collected at \(f_s=1000\text{ Hz}\). Find the DFT frequency-bin spacing.

Practice 8: Nyquist Frequency

A signal is sampled at \(8000\text{ samples/s}\). Find the Nyquist frequency.

Practice 9: Scaling

If

find the transform of \(x(4t)\).

Practice 10: Choose the More Useful Domain

For each task, identify whether the time domain or frequency domain is usually the more direct first representation:

1. measuring the arrival time of a pulse;
2. finding harmonic content in a periodic waveform;
3. designing a filter to remove a narrow interference band;
4. measuring settling time after a step input.

Practice Answers

Practice 1

Use

Therefore,

Practice 2

The fundamental frequency is

Thus,

Practice 3

A delay by \(3\) produces a phase factor:

Practice 4

Modulation shifts the spectrum:

Practice 5

Two derivatives produce

Therefore,

Practice 6

For an LTI system,

Hence,

Practice 7

The frequency spacing is

Therefore,

Practice 8

The Nyquist frequency is half the sampling rate:

Practice 9

By the scaling property,

Practice 10

1. Time domain: arrival time is directly localized in time.
2. Frequency domain: harmonic components appear directly as spectral structure.
3. Frequency domain: the unwanted band can be identified and attenuated spectrally.
4. Time domain: settling time is measured from the evolving response.

What to Remember

The central transform relationship is

• the time domain shows when events occur and how signals evolve;
• the frequency domain shows how oscillatory components are distributed;
• magnitude and phase are both essential parts of a complex spectrum;
• periodic signals naturally produce discrete harmonic frequencies;
• aperiodic signals generally require continuous spectra;
• short duration and narrow bandwidth cannot generally be achieved simultaneously;
• time shifts become phase factors;
• modulation shifts spectra;
• differentiation becomes multiplication by \(i\omega\);
• convolution in time becomes multiplication in frequency;
• an LTI system is characterized spectrally by its frequency response \(H(\omega)\);
• sampling introduces a Nyquist limit and the possibility of aliasing;
• finite observation and windowing affect spectral estimates;
• the best domain is the one that makes the structure of the problem simplest.

The most important habit is to move deliberately between representations rather than treating one domain as the default for every problem.

Where We Go Next

Frequency-domain analysis works so well for linear time-invariant systems because complex exponentials are preserved in form: the system changes only their coefficients. That property is not unique to Fourier analysis. It is an example of a much broader mathematical idea involving eigenfunctions.

The next lesson, Eigenfunction Expansions, generalizes the spectral viewpoint beyond ordinary Fourier frequency. We will study how functions can be expanded in eigenfunctions of differential and integral operators, how boundary conditions generate discrete spectra, and why choosing operator-adapted modes can turn difficult equations into simpler coefficient problems.