Wavelets: a mathematical microscope

Wavelet transform is presented as a “mathematical microscope” for signals that are noisy, irregular, and structured at multiple time scales—letting analysts quantify what’s happening when, not just what frequencies exist. The core problem is that Fourier transform, while powerful at decomposing a signal into frequency components, collapses time information: it can tell which frequencies are present but not when they occur. That limitation becomes critical in real-world scenarios where patterns appear, evolve, and disappear.

The explanation starts with time–frequency duality using a simple two-number example: sending the sum and difference of two values preserves all information, just in a different representation. Extending that idea to signals leads to Fourier transform, which decomposes a function into a sum of sines and cosines. Each frequency’s contribution can be read from the frequency domain, and an inverse transform can reconstruct the original signal. But Fourier’s single-resolution trade-off is unavoidable: perfect time resolution and perfect frequency resolution cannot coexist. The traffic-light analogy makes the point. Whether the colors occur in a fixed order, a scrambled order, or even overlap simultaneously, the Fourier transform still shows the same frequency peaks—because it is blind to temporal dynamics.

Wavelets enter as the compromise tool. Instead of using sine waves that extend infinitely in time, wavelet transform uses localized “little waves” that are confined to finite time windows. A wavelet is defined as a family of functions (not a single one) that satisfy two key constraints: zero mean (no zero-frequency/average component) and finite energy (the integral of the squared function over all time is finite), which enforces time localization. The Morlet wavelet is highlighted as a common choice: a cosine (or complex exponential) damped by a Gaussian envelope.

The mechanics of wavelet transform are built step-by-step. A mother wavelet is modified by two parameters: translation in time (shifting by b) and scaling (stretching or compressing by a, equivalently changing frequency). For each (a, b), the method computes how well the scaled-and-shifted wavelet matches the signal locally. That “match” is expressed as an integral of the product of the signal and the wavelet—mathematically equivalent to a dot product, capturing similarity through sign-aligned agreement. Sliding the wavelet across time resembles convolution, producing a short-lived oscillatory response when the signal’s local frequency aligns with the wavelet’s.

To convert that response into a meaningful measure of frequency power over time, the transform uses the complex form of the Morlet wavelet. Convolution is computed with both real and imaginary parts; the magnitude (absolute value) of the resulting complex coefficient gives the intensity of a frequency component at each time. Plotting magnitude across scales and times yields a wavelet scalogram, which makes evolving frequency content visible—such as a sine wave whose frequency increases over time, or brain rhythms that show distinct bouts of low-frequency activity with higher-frequency components riding on top.

Finally, wavelet transform is framed as respecting the uncertainty principle rather than violating it. The time–frequency plane is described using “Heisenberg boxes”: low frequencies get better frequency precision but poorer time precision (wide, short boxes), while high frequencies get better time precision but poorer frequency precision (tall, narrow boxes). The result is an optimal, frequency-dependent balance that reveals when patterns begin and end—exactly the information Fourier transform tends to miss.