Cross-correlation between two signals measures their similarity as a function of relative time delay. The peak of the cross-correlation function indicates the delay of maximum similarity, enabling time-delay estimation and synchronization. Normalized cross-correlation (correlation coefficient) is independent of signal amplitudes. Applications include radar/sonar target detection, audio alignment, and template matching.
Cross-correlate a known template with a signal containing the template at unknown delay. Find the delay by locating the correlation peak. Add noise and observe robustness.
From your study of convolution, you know how to compute the output of a linear system when you know its impulse response: slide the input across the impulse response, multiply, and sum at each lag. Cross-correlation is the same sliding-and-summing operation, but with a different goal — instead of computing a system output, you are measuring *similarity between two signals as a function of relative time shift*. Formally, R_xy(τ) = ∫ x(t) y(t + τ) dt for continuous signals. The key difference from convolution is that one signal is *not* time-reversed before sliding. Where convolution asks "how does a system respond to this input?", cross-correlation asks "at what delay does signal y most resemble signal x?"
The core application is time delay estimation. Suppose a sonar pulse travels from a transmitter, reflects off an underwater object, and arrives at two hydrophones spaced some distance apart. The reflected signal arrives at the nearer hydrophone first, then at the farther one. If you cross-correlate the two hydrophone recordings, the result will be a function that peaks at the lag equal to the travel time difference between the two paths. That peak lag, multiplied by the speed of sound in water, gives you the difference in path lengths — enough to triangulate the target's position. This is the operating principle behind sonar, GPS multipath analysis, seismic source location, and audio source localization in room acoustics.
Normalized cross-correlation divides by the product of the signals' energies (or standard deviations), producing values between −1 and +1. This matters when comparing signals of different amplitudes: the unnormalized peak could be large simply because one signal has high energy, not because they are particularly similar. The normalized version tells you about *shape* similarity rather than magnitude. In template matching — finding a small image patch within a larger image, or detecting a known waveform in noisy data — normalized cross-correlation is the standard tool because it is insensitive to illumination changes or amplitude variations in the target.
One subtlety worth internalizing: cross-correlation and convolution are related by time-reversal of one signal. R_xy(τ) = x(−τ) * y(τ), where * denotes convolution. This means all the computational machinery you know for convolution — the convolution theorem in the frequency domain, FFT-based fast computation — applies directly. In practice, cross-correlating long signals is always done via the FFT: transform both signals, multiply X*(ω)·Y(ω) (the conjugate of one by the other), then inverse transform. This reduces an O(N²) sliding-dot-product computation to O(N log N), making correlation of long audio recordings, long sensor streams, or large images computationally feasible. The correlation peak may be sharpened further with whitening (pre-filtering to flatten the power spectrum), which is valuable when the shared signal has a narrow spectral peak that would otherwise spread the correlation lobe.