Sampling, Aliasing, and Reconstruction

How continuous-time signals become discrete-time data, why under-sampling causes aliasing, and when bandlimited signals can be reconstructed.
Modified

April 26, 2026

Keywords

sampling, aliasing, Nyquist, reconstruction, bandlimited signal

1 Role

This is the third page of the Signal Processing and Estimation module.

Its job is to explain how continuous-time signals become discrete-time data without losing the frequency story built on the previous page.

The first two pages said:

  • LTI systems are described by convolution
  • frequency response explains systems spectrally

This page adds:

  • sampling copies spectra
  • under-sampling causes overlap
  • bandlimiting plus enough sampling makes reconstruction possible

2 First-Pass Promise

Read this page after Fourier Analysis, Frequency Response, and Spectral Views.

If you stop here, you should still understand:

  • why sampling is a modeling step, not only a hardware detail
  • what the Nyquist idea means at first pass
  • what aliasing is and why it happens
  • why reconstruction needs both a bandlimit story and an anti-aliasing story

3 Why It Matters

Most modern systems live in a continuous-discrete loop.

The world often generates continuous-time signals:

  • sound pressure
  • voltage traces
  • motion trajectories
  • optical intensity

But computers, sensors, and storage systems work with samples.

So signal processing needs a language for:

  • converting continuous-time signals into discrete-time data
  • knowing when that conversion loses information
  • reconstructing a continuous-time signal from samples when conditions are favorable

This matters in communication, sensing, imaging, and ML because aliasing is not just a classroom artifact.

It is a real failure mode whenever high-frequency content is hidden inside low-rate observations.

4 Prerequisite Recall

  • Fourier representations describe signals by their frequency components
  • the spectrum tells us where oscillatory content lives
  • LTI filtering can suppress or preserve different frequency bands
  • low-pass behavior is often the right first-pass model for smoothing or anti-alias filtering

5 Intuition

5.1 Sampling Turns A Continuous Signal Into A Sequence

Sampling means measuring a continuous-time signal at regularly spaced times:

\[ x[n] = x_c(nT), \]

where T is the sampling period and f_s = 1/T is the sampling rate.

So a continuous waveform becomes a discrete sequence.

5.2 Sampling Repeats The Spectrum

The spectral viewpoint says sampling is not neutral.

At first pass, sampling takes the original spectrum and creates shifted copies spaced by the sampling frequency.

If those copies stay separated, the original signal can be recovered.

If they overlap, information is mixed together.

That overlap is aliasing.

5.3 Aliasing Means Different Frequencies Become Indistinguishable

When the sampling rate is too low, a high-frequency sinusoid can produce exactly the same samples as a lower-frequency sinusoid.

So after sampling, the sequence no longer tells us which frequency was really present in continuous time.

That is why aliasing is a loss-of-identifiability problem, not just a visual artifact.

5.4 Nyquist Is A Separation Condition

If a signal is bandlimited and the sampling rate is at least twice the highest frequency present, then the repeated spectra stay separated.

At first pass, this is the Nyquist idea:

  • enough spacing in time means enough separation in frequency copies

5.5 Reconstruction Needs Anti-Aliasing

Real signals are often not perfectly bandlimited.

So practical systems usually low-pass filter before sampling to remove frequency content that would fold back and alias.

That pre-filter is the anti-aliasing filter.

Then a reconstruction filter can rebuild a continuous-time signal from the clean enough sampled sequence.

6 Formal Core

Definition 1 (Definition: Uniform Sampling) Uniform sampling of a continuous-time signal x_c(t) with period T produces the discrete-time sequence

\[ x[n] = x_c(nT). \]

The sampling rate is

\[ f_s = \frac{1}{T}. \]

Definition 2 (Definition: Bandlimited Signal) A continuous-time signal is bandlimited to bandwidth W if its spectrum is zero outside the frequency range |\omega| \le W or, equivalently, outside |f| \le B in ordinary-frequency notation.

At first pass, bandlimited means:

  • there is a highest frequency that matters

Theorem 1 (Theorem Idea: Sampling Replicates The Spectrum) Uniform sampling produces shifted copies of the original continuous-time spectrum, spaced by the sampling frequency.

This is the main spectral fact behind both reconstruction and aliasing.

Theorem 2 (Theorem Idea: Nyquist Sampling Condition) If a signal is bandlimited and the sampling rate is at least twice the highest frequency present, then ideal reconstruction from uniform samples is possible.

At first pass, this is often stated as:

\[ f_s \ge 2B. \]

The rate 2B is the Nyquist rate.

Definition 3 (Definition: Aliasing) Aliasing occurs when the spectral copies created by sampling overlap, making distinct continuous-time frequency components indistinguishable in the sampled sequence.

Theorem 3 (Theorem Idea: Anti-Aliasing Before Sampling) If a signal contains frequency content above what the sampling rate can safely represent, low-pass filtering before sampling can prevent those components from folding into lower frequencies.

This is why anti-alias filtering is a structural part of sampling pipelines.

7 Worked Example

Take the continuous-time signals

\[ x_1(t)=\sin(2\pi \cdot 1\, t), \qquad x_2(t)=\sin(2\pi \cdot 9\, t). \]

Now sample at f_s = 8 Hz, so T = 1/8.

Then

\[ x_1[n] = \sin\!\left(2\pi \cdot \frac{1}{8} n\right) \]

and

\[ x_2[n] = \sin\!\left(2\pi \cdot \frac{9}{8} n\right) = \sin\!\left(2\pi n + 2\pi \cdot \frac{1}{8} n\right) = \sin\!\left(2\pi \cdot \frac{1}{8} n\right). \]

So the 1 Hz signal and the 9 Hz signal produce the same samples.

That is aliasing in its cleanest possible form:

  • two different continuous-time frequencies
  • one identical discrete-time sequence

So no downstream algorithm can recover which original frequency was present from the samples alone.

8 Computation Lens

When a pipeline moves between continuous time and discrete time, ask:

  1. what is the actual sampling rate?
  2. what frequency content is present before sampling?
  3. is there enough bandlimit structure for alias-free representation?
  4. is an anti-alias filter present before sampling?
  5. is the reconstruction claim exact, approximate, or only good enough for the application?

These questions matter in data collection, simulation, audio, imaging, sensor fusion, and sequence modeling.

9 Application Lens

9.1 Audio And Imaging

Aliasing shows up as audible distortion in sound and as moire or jagged artifacts in images and video.

9.2 Communication And Sensing

Sampling determines what part of an analog waveform a digital receiver can even hope to recover.

9.3 ML Pipelines

Many learned systems consume sampled signals, images, spectrograms, or time series whose quality already depends on earlier sampling and anti-aliasing choices.

10 Stop Here For First Pass

If you stop here, retain these five ideas:

  • sampling converts continuous-time signals into discrete-time sequences
  • sampling creates repeated copies of the spectrum
  • aliasing is spectral overlap caused by sampling too slowly
  • bandlimited signals can be reconstructed when the sampling rate is high enough
  • anti-alias filtering matters before sampling, not after the damage is already done

11 Go Deeper

The strongest next page is:

The strongest adjacent live pages are:

12 Optional Deeper Reading After First Pass

13 Sources and Further Reading

Back to top