Questions: Cross-Correlation and Time Delay Estimation

Question 1

A radar system transmits pulse x(t) and receives y(t) = x(t − τ₀) + noise, where τ₀ is the unknown round-trip travel time to a target. The cross-correlation Rxy(τ) is computed. How is target range estimated from this result?

• A. By measuring the amplitude of Rxy(τ) at τ = 0, which represents the signal energy remaining after transmission loss over distance.
• B. By finding the lag τ where Rxy(τ) is maximized — this equals the round-trip delay τ₀, and multiplying by wave propagation speed gives range.
• C. By subtracting the noise floor from Rxy(τ) and dividing by the signal bandwidth to extract timing information.
• D. By integrating Rxy(τ) over all lags, since the integral equals the total signal power, which decays with distance.

Answer: (B)

Cross-correlation measures similarity between x and a time-shifted version of y. When y is a delayed copy of x (plus noise), Rxy(τ) peaks at the lag τ = τ₀ that best aligns x with the delayed signal — exactly the round-trip delay. The amplitude at τ = 0 gives only autocorrelation of x with the unshifted y, which is irrelevant to delay estimation. Range = (τ₀/2) × wave_speed (dividing by 2 because the delay is round-trip). This is the operating principle of all delay-and-estimate systems: radar, sonar, GPS pseudorange, and ultrasonic flow meters.

Question 2

Signal x(t) has amplitude 1 and signal y(t) is identical in shape to x(t) but has amplitude 10 and a 5-second time lag. You compute both the raw cross-correlation Rxy(τ) and the normalized cross-correlation ρxy(τ). Which statement correctly describes what normalization reveals that the raw cross-correlation does not?

• A. Normalization reveals the time lag more precisely by suppressing noise at lags far from the peak.
• B. Normalization confirms that the shape similarity is essentially perfect (ρ ≈ +1 at τ = 5 s) regardless of the amplitude difference — isolating shape similarity from amplitude effects.
• C. Normalization reveals the frequency content of the delay, showing which spectral components are responsible for the lag.
• D. The two are equivalent; normalization only rescales the vertical axis without changing any interpretation.

Answer: (B)

The raw Rxy(τ) peak height depends on both signal amplitudes and shape similarity — a high peak could mean either 'very similar shapes' or 'very large amplitudes' or both. Normalization by √(Rxx(0)·Ryy(0)) scales the result to [−1, +1], removing amplitude dependence entirely. A peak of ρ ≈ +1 means the signals are essentially identical in shape up to the given delay, regardless of how large or small either signal is. This is critical for pattern matching: finding a known template in a noisy signal requires responding to shape, not amplitude. Option D is wrong because normalization changes the interpretation of the peak height, even though it does not change the peak location.

Question 3

The cross-correlation Rxy(τ) usually peaks at τ = 0, just as the autocorrelation Rxx(τ) does.

• T. True
• F. False

Answer: False

The autocorrelation Rxx(τ) peaks at τ = 0 because a signal is always most similar to an unshifted version of itself — any shift reduces overlap. Cross-correlation Rxy(τ) peaks at the lag τ₀ that best aligns signal x with signal y. If y is a delayed version of x (e.g., an echo), the peak is at τ = τ₀, not zero. If x and y have no time delay relationship and are uncorrelated, there may be no sharp peak at all. The cross-correlation peak location is information — it reveals the relative time alignment between two signals, which is exactly what makes it useful for delay estimation.

Question 4

Computing cross-correlation efficiently using the FFT requires only a forward transform of each signal, a pointwise complex multiplication of their spectra, and an inverse transform to obtain the correlation in the time domain.

• T. True
• F. False

Answer: True

The cross-correlation theorem states that Rxy(τ) ↔ X*(f)·Y(f), where X*(f) is the complex conjugate of X(f). This means the Fourier transform of the cross-correlation equals the cross-spectrum. So to compute Rxy: (1) FFT x to get X(f), (2) FFT y to get Y(f), (3) multiply X*(f)·Y(f) pointwise, (4) inverse FFT the result. This is O(N log N) rather than O(N²) for direct computation of the correlation integral — a crucial practical speedup for long signals. This same FFT-based approach underlies matched filtering and generalized cross-correlation with frequency-domain weighting.

Question 5

Explain in your own words what the normalized cross-correlation ρxy(τ) reveals, and why normalization matters when comparing signals of very different amplitudes.

Model answer: The normalized cross-correlation ρxy(τ) = Rxy(τ) / √(Rxx(0)·Ryy(0)) measures the degree of linear similarity between x and y at each time lag, bounded between −1 and +1. A value of +1 at some lag means the signals are essentially identical in shape (one is a scaled version of the other at that delay); −1 means they are inverted copies; 0 means they are uncorrelated. Normalization removes amplitude dependence: a loud signal and a quiet one with identical shape will give ρ ≈ +1 at the correct lag, just as two signals of equal amplitude would. Without normalization, the raw peak height conflates shape similarity with signal energy — you cannot tell whether a large Rxy value means 'similar shapes' or 'large amplitudes' or both. Normalization isolates the shape and delay information, which is what detection and estimation tasks actually need.

Think of normalization as measuring 'how much of signal y can be explained by a shifted, scaled copy of signal x.' A value of +1 says 'perfectly, up to a scalar.' The division by signal energies is the same operation as computing a correlation coefficient in statistics — it standardizes the result to be interpretable as a measure of dependence rather than magnitude.