Does it account for correlations?

Yes, through the inverse covariance matrix.

What happens with one variable?

The classical one-sample statistic reduces to the ordinary t statistic squared.

Does rejection mean all variables differ?

No. It rejects the specified mean vector jointly.

AI Glossary / Statistics and probability / Intermediate

Hotelling's t-squared statistic

Definition

Hotelling's t-squared statistic, conventionally written T-squared or T^2, measures a multivariate departure while accounting for covariance among variables. Its classical one-sample form tests whether a population mean vector equals a specified target vector. It extends the idea behind the squared Student t statistic to several measurements considered jointly, with related forms used for comparing two groups and monitoring processes.

What it means

A business often cares about a combination of outcomes rather than one number. Delivery time and processing cost, for example, may move together. Judging each separately can miss an unusual combination or create a multiple-testing problem.
T-squared measures departure in the geometry defined by the covariance matrix. A change in a normally variable direction contributes less than a comparable change in a stable direction. The statistic is a joint discrepancy measure, not an explanation of what caused the difference.

How it works

For a classical one-sample test, observe n independent p-dimensional measurement vectors. Calculate their sample mean vector x-bar and sample covariance matrix S, using divisor n minus 1. Specify the target mean vector mu-zero before examining the test result.
Let d equal x-bar minus mu-zero. The statistic is T^2 = n*d-transpose*S-inverse*d. The transpose turns the first vector into a row so the matrix product produces one nonnegative number when S is positive definite.
Covariances matter because they describe typical combinations of movement. Two positively correlated measurements rising together may be less unusual than one rising while the other falls. Merely adding separately standardized squared differences ignores this structure.
Under independent observations from a multivariate normal population with a positive-definite covariance matrix, and with n greater than p, the null-reference statistic is F = (n - p)*T^2/[p*(n - 1)]. It follows an F distribution with p and n minus p degrees of freedom under the null hypothesis.
These finite-sample assumptions matter. Outliers, dependence, or substantial departures from the assumed distribution can undermine interpretation. Large-sample approximations may be available, but they require their own justification rather than automatic acceptance.
When p is at least n, the ordinary sample covariance matrix cannot have full rank. Even when an inverse exists, strong collinearity can make calculations unstable. Dimension reduction or regularization changes the procedure and generally requires different calibration of the test.
For two independent groups, a classical version uses their difference in mean vectors and a pooled covariance estimate, with equal population covariances assumed. Paired comparisons instead apply a one-sample procedure to vectors of within-pair differences. These are distinct designs with distinct formulas.
Related T-squared control charts examine multivariate process departures. Their limits depend on how reference parameters were estimated and whether historical or future observations are monitored. Do not transplant a one-sample mean-test threshold directly into every chart.
A significant joint result should prompt examination of effect sizes, uncertainty, and suitable follow-up analyses. It does not establish that every measured variable changed, and unadjusted follow-up tests can create additional false-positive risk.

Three practical examples

An AI operations team checks whether a joint mean vector of latency and processing cost differs from a specified operational target.
A quality team compares several correlated product measurements across two production settings using a design-appropriate multivariate test.
A process-monitoring team tracks combined measurement departures with a T-squared chart whose limits match its reference-data procedure.

Think of it this way

Think of a tilted target drawn around a cloud of observations. The target stretches along directions where the measurements usually vary together. T-squared measures how far a result lies from the center after allowing for that shape, rather than judging horizontal and vertical distances independently.

Worked example

Suppose n = 10 and p = 2. The observed mean differs from the target by d = (1, 2). Assume the sample covariance matrix is S = [[4, 2], [2, 9]]. Its determinant is 4*9 - 2*2 = 32, so S-inverse = [[9, -2], [-2, 4]]/32.
The quadratic form is (9*1^2 - 4*1*2 + 4*2^2)/32 = 17/32. Therefore T^2 = 10*17/32 = 85/16 = 5.3125.
The transformed statistic is F = 8*5.3125/(2*9) = 85/36, approximately 2.36111, with 2 and 8 degrees of freedom under the classical assumptions. Its upper-tail probability is approximately 0.15635. At a preselected 0.05 level, this example would not reject the target mean vector.
That decision does not prove equality or operational acceptability. It concerns evidence under a stated sampling model, not whether the observed departures are small enough for the business.

Illustrative case study

In this fictional case, Harbour Automation evaluates average response latency and processing cost together. Its analyst initially considers two separate tests, then notes that the outcomes are correlated and uses a justified joint mean test. The report includes the two observed mean differences and their practical implications. Management distinguishes failure to reject the statistical target from demonstrating that both metrics meet contractual tolerances.

Common misconceptions and mistakes

T-squared is not an ordinary Euclidean distance and is not generally the sum of separate squared t statistics. A significant result does not identify the responsible variable or establish causation. A non-significant result does not prove equivalence.
The same-looking quadratic form can have different scaling and reference distributions in mean testing, individual anomaly scoring, and control charts. An invertible covariance estimate alone does not establish the assumptions needed for a valid p-value.

Business use

Use the statistic when a joint mean comparison matches the decision and sufficient data support covariance estimation. Report the measurement definitions, sampling design, mean departures, assumptions, and follow-up strategy. For high-dimensional AI features, check rank and stability before treating a classical T-squared calculation as a calibrated test.

Check your understanding

Why is covariance included?

Answer: It distinguishes usual combinations of movement from unusual ones.

Can the classical formula be used unchanged when p exceeds n?

Answer: No. The ordinary sample covariance is singular, and a different method with justified inference is needed.

Related terms

Mahalanobis distance
Covariance matrix
F distribution
Multivariate analysis
Student t-test
multivariate control chart

Learn next

Study covariance geometry and one-sample multivariate mean tests. Then learn simultaneous confidence regions, two-sample designs, and the distinction between inference about means and monitoring individual observations.

Sources

NIST/SEMATECH, Hotelling's T squared: https://www.itl.nist.gov/div898/handbook/pmc/section5/pmc543.htm NIST/SEMATECH, Hotelling Control Charts: https://www.itl.nist.gov/div898/handbook/pmc/section3/pmc341.htm Penn State STAT 505, Inferences Regarding Multivariate Population Mean: https://online.stat.psu.edu/stat505/Lesson07

Previous: Hot languages Next: Hough transform