Can entries exceed one?

Yes, under the uncentered cross-moment convention.

When does cross-correlation equal cross-covariance?

When the mean outer product is zero, including when both vectors are zero-mean.

Does a rectangular matrix indicate an error?

No. Different vector lengths naturally produce a rectangular matrix.

AI Glossary / Statistics and probability / Advanced

Cross-correlation matrix

Definition

A cross-correlation matrix describes pairwise relationships between components of two random vectors. In the probability and signal-processing convention used here, it is the matrix R_XY = E[XY transpose], containing uncentered cross-moments. Other fields use the phrase for matrices of normalized Pearson correlations, so the convention must always be stated.

What it means

Let X contain measurements from one group of sensors and Y contain measurements from another. A cross-correlation matrix puts each X component against every Y component, producing one entry per pair.

Under the uncentered convention, entry i,j is E[X_iY_j]. It includes the contribution of the means and is not generally bounded between -1 and 1. The cross-covariance matrix instead centers both vectors, while a matrix of Pearson correlations also divides by their standard deviations.

These distinctions are essential when reading documentation. A matrix of time-lagged signal cross-correlations adds another choice: which lag and normalization define each displayed value. The same short name can therefore refer to meaningfully different objects.

How it works

For real vectors X of length m and Y of length n, take their outer product XY transpose and average over the joint distribution. The result has m rows and n columns. It need not be square or symmetric.

Reversing the vectors gives R_YX = R_XY transpose. With means mu_X and mu_Y, cross-covariance is C_XY = R_XY - mu_X mu_Y transpose. If both vectors have zero means, these two matrices coincide.

For paired observations, an empirical uncentered estimate averages their outer products. Centering produces a covariance estimate whose denominator convention must be specified. Normalizing covariance by appropriate positive standard deviations produces Pearson coefficients. Preserve pairing, variable order, units, and any lag definition throughout the calculation.

Three practical examples

A sensor-fusion system summarizes cross-moments between two sets of simultaneous measurements before constructing a linear estimator.
A signal analyst compares multichannel inputs and outputs at a specified delay, documenting whether the calculation centers or normalizes the signals.
A machine-learning researcher examines a normalized cross-correlation block between two feature representations, explicitly distinguishing it from an uncentered second-moment matrix.

Think of it this way

Think of a rectangular comparison board: one set of variables labels the rows and another labels the columns. Each cell compares one row variable with one column variable. The board's meaning depends on the comparison rule; raw products, centered products, and standardized products are different rules.

Worked example

Consider an original joint distribution with two equally likely states. In state one, X = [1,0] and Y = [2,1,0]. In state two, X = [3,2] and Y = [0,3,4]. Treat these as population states with probability one-half each.

The first outer product is [[2,1,0],[0,0,0]]. The second is [[0,9,12],[0,6,8]]. Their probability-weighted average gives R_XY = [[1,5,6],[0,3,4]].

The means are mu_X = [2,1] and mu_Y = [1,2,2]. Their outer product is [[2,4,4],[1,2,2]]. Subtracting it gives C_XY = [[-1,1,2],[-1,1,2]].

Notice that the uncentered entry relating X_1 and Y_1 is positive, 1, while the centered covariance is negative, -1. The mean contribution changes the interpretation.

The cross-correlation matrix is 2-by-3 and therefore cannot be symmetric. Reversing the two vectors produces its 3-by-2 transpose. Values such as 6 are valid uncentered cross-moments; they are not impossible Pearson correlations. No sample degrees-of-freedom correction is used because this example specifies the entire two-state distribution.

Illustrative case study

Fictional case: Vale Robotics receives a cross-correlation matrix from a supplier and compares it with an internal analytics report. The numbers disagree, and some supplier values exceed one.

The engineers inspect the definitions before assuming a software error. The supplier reports uncentered cross-moments, while the internal report centers and standardizes the measurements. They also find that one report reverses the order of the sensor groups.

After aligning the convention, units, and indexing, the team can compare the outputs meaningfully. It retains both representations where useful, documenting why the estimator requires one and the diagnostic dashboard uses the other. The case illustrates a definition mismatch rather than a real supplier incident.

Common misconceptions and mistakes

An uncentered cross-correlation matrix is not necessarily a matrix of Pearson coefficients.
Cross-correlation between distinct vectors need not be square, symmetric, or positive semidefinite by itself.
The special autocorrelation matrix E[XX transpose] is symmetric positive semidefinite for real X; that property should not be copied indiscriminately to a cross-block.
Independence gives R_XY = mu_X mu_Y transpose, not necessarily a zero matrix.
Zero cross-covariance does not generally establish independence.
In time-series work, reversing signal order can reverse the interpretation of lag as well as transpose channel relationships.

Business use

The matrix is useful when integrating measurements, comparing representations, or estimating relationships between inputs and outputs. Business and engineering teams should request the exact formula rather than relying on a label such as correlation report.

For empirical estimates, confirm that rows represent synchronized or otherwise correctly matched observations. Mixing timestamps can create a misleading picture, particularly when signals trend or share periodic behavior. Missing measurements and preprocessing can further change the available pairs.

Separate descriptive relationships from causal claims. A large entry can reflect units, nonzero means, a common driver, or genuine co-movement. Decide which quantity answers the operational question before ranking entries or using them in an automated rule.

Check your understanding

Answer: 1.

What is the shape when X has two components and Y has three?

Answer: Two rows by three columns. 2.

Why did one example entry change sign after centering?

Answer: Its positive mean-product contribution was removed. 3.

What must be checked before comparing two reports?

Answer: Their formulas, centering, normalization, variable order, units, and any lag conventions.

Related terms

Correlation matrix
Cross-covariance matrix
outer product
second moment
random vector
signal cross-correlation
sensor fusion

Learn next

Learn covariance and outer products, then compare uncentered and centered matrix calculations on the same data. Study normalized correlations and lagged signal relationships separately. Practice checking dimensions before applying matrix formulas in a model.

Sources

Hossein Pishro-Nik, Introduction to Probability, Statistics, and Random Processes, Random Vectors: https://www.probabilitycourse.com/chapter6/6_1_5_random_vectors.php

Eindhoven University of Technology, SPS Education, Random vectors: https://spseducation.tue.nl/courses/5cta0/mathematicalbackground_probability_vectors/

Primary educational definitions were consulted. The entry explicitly adopts the uncentered real-vector convention and distinguishes normalized usage. The two-state calculation and fictional case are original.

Previous: Cross-brackets Next: Cross decomposition