Harbor 🪼

6. Fourier Transform

4 min read

The Fourier Transform (FT) is a mathematical bridge that allows us to view data from two different perspectives: the Time/Spatial Domain (how a signal changes over time or space) and the Frequency Domain (what frequencies make up that signal).

The “Smoothie” Intuition

A common way to understand the Fourier Transform is the Smoothie Analogy:

  • The Signal (Time Domain): This is the smoothie itself. You can taste the final result, but it’s hard to tell exactly how much of each ingredient went into it just by looking at it.
  • The Fourier Transform: This is the “de-blender” or recipe generator. It analyzes the smoothie and tells you: “This is 40% banana, 30% strawberry, and 30% yogurt.” It breaks a complex, messy signal into its fundamental components (frequencies).
  • The Inverse Fourier Transform: This is the blender. It takes the individual ingredients (frequencies) and combines them to recreate the original signal.

History

  • Jean-Baptiste Joseph Fourier (1807): A French mathematician who proposed that any periodic function could be represented as a sum of sines and cosines. He discovered this while studying how heat flows through metal rods. At the time, famous mathematicians like Lagrange and Laplace were skeptical.
  • The FFT Revolution (1965): While the math existed, calculating it was slow (). James Cooley and John Tukey published the Fast Fourier Transform (FFT) algorithm, which reduced the complexity to . This algorithm is widely considered one of the most important in history, as it made real-time digital signal processing possible.

Mathematical Definition

Continuous Fourier Transform

For a continuous signal :

  • represents the frequency.
  • is the “probe” (a complex sinusoid) we use to check for the presence of frequency in the signal.

Discrete Fourier Transform (DFT)

In computers, we work with discrete samples : This is what the FFT computes efficiently.

Core Properties

  1. Linearity: The transform of a sum of signals is the sum of their individual transforms.
  2. Frequency Shifting: Shifting a signal in time results in a phase shift in the frequency domain.
  3. Convolution Theorem (Crucial for ML): Convolution in the time domain is equivalent to point-wise multiplication in the frequency domain: This is how deep learning libraries (like PyTorch or TensorFlow) speed up convolutions in CNNs.

Applications

1. Signal & Image Processing

  • Filtering: To remove high-frequency noise (like a hiss in audio), you transform to the frequency domain, zero out the high frequencies, and transform back.
  • Compression: JPEG and MP3 formats work by discarding frequencies that the human eye or ear cannot easily perceive. This is why a high-quality image can be compressed into a small file.

2. Machine Learning

  • Feature Extraction: For audio data (speech recognition) or time-series data (stock market, sensor logs), we often convert the raw signal into a Spectrogram (a visual representation of the frequency spectrum over time). This makes it much easier for models like LSTMs or Transformers to find patterns.
  • Accelerating CNNs: Because of the Convolution Theorem, very large convolution kernels can be computed much faster in the frequency domain using the FFT.
  • Anomaly Detection: In industrial IoT, Fourier analysis can detect “rhythms” in machine vibrations. If the frequency spectrum suddenly changes, it’s a sign that a part is about to fail.

3. Science & Engineering

  • Astronomy: Analyzing the light from stars (spectroscopy) to determine their chemical composition.
  • Quantum Mechanics: The Heisenberg Uncertainty Principle is essentially a property of the Fourier Transform (you cannot be precisely “located” in both time and frequency at once).

Backlinks