Why Positive Semidefiniteness and Rank Matter
A covariance matrix cannot assign negative variance to any linear combination of the quantities it describes. This basic fact places a structural constraint on the covariance matrix of shared-control comparisons. It also lets us determine whether the matrix is invertible, whether some contrast is redundant, and whether a proposed set of covariance entries could have arisen from the stated sampling model.
The introductory material on building the shared-control covariance matrix gives the entries of that matrix. Here, the focus is on what those entries imply as a whole. In the usual setup with one distinct treatment-control comparison for each treatment arm, the matrix is not merely positive semidefinite: under nondegenerate independent treatment-arm sampling, it is positive definite and has full rank.
Model and Shared-Control Covariance Matrix
Suppose there are \(m\) treatment arms and one shared control arm. For this tutorial, assume independent sampling across arms, a common outcome variance \(\sigma^2>0\), and finite positive sample sizes \(n_0,n_1,\ldots,n_m\). Write \(\bar{X}_i\) for the sample mean in arm \(i\), with arm \(0\) denoting control. Thus, \(\operatorname{Var}(\bar{X}_i)=\sigma^2/n_i\), and sample means from different arms are independent. No normality assumption is needed for the covariance and rank results below.
For treatment \(i\), define its treatment-control contrast as \(D_i=\bar{X}_i-\bar{X}_0\). Each contrast contains the same control mean. For distinct treatments \(i\) and \(j\), the treatment-arm means are independent of each other and of the control mean, so the only shared random component is \(\bar{X}_0\). The resulting covariance matrix for \(D=(D_1,\ldots,D_m)^\mathsf{T}\) is
where \(\mathbf{1}\) is the \(m\)-vector of ones. The diagonal entries are \(\sigma^2(1/n_i+1/n_0)\); every off-diagonal entry is \(\sigma^2/n_0\). The formula follows by expanding \(\operatorname{Cov}(\bar{X}_i-\bar{X}_0,\bar{X}_j-\bar{X}_0)\), using independence across arms, and retaining the control variance shared by both contrasts.
Quadratic-Form Proof and Rank
To establish the result, take any real coefficient vector \(a=(a_1,\ldots,a_m)^\mathsf{T}\). The quadratic form \(a^\mathsf{T}\Sigma a\) is the variance of the linear combination \(\sum_{i=1}^m a_iD_i\). Substituting the matrix formula gives
Both terms inside the parentheses are nonnegative, so the quadratic form is nonnegative for every \(a\). That proves positive semidefiniteness directly. If \(a\neq 0\), at least one \(a_i\neq 0\). Because \(n_i\) is finite and positive, \(\sum_i a_i^2/n_i>0\). Since \(\sigma^2>0\), the whole quadratic form is strictly positive. Thus \(\Sigma\) is positive definite. A positive-definite \(m\times m\) matrix has no zero eigenvalues, so its rank is \(m\).
This proof also explains why the shared control does not reduce the rank in the standard setup. The control component contributes a variance term involving \((\sum_i a_i)^2\), but each treatment contrast also contains its own treatment-arm sampling noise. For every nonzero coefficient vector, that treatment-specific contribution is strictly positive. The rank conclusion depends on having positive sampling variance in each distinct treatment arm; it is not a conclusion based on the shared-control term alone.
Worked Example: Checking Positive Definiteness With Two Treatments
Consider a hypothetical trial with two treatment arms and one control arm. Let \(\sigma^2=16\), \(n_0=16\), \(n_1=8\), and \(n_2=16\). The two contrast variances are \(16(1/8+1/16)=3\) and \(16(1/16+1/16)=2\); their covariance is \(16/16=1\). Therefore,
For an arbitrary \(a=(a_1,a_2)^\mathsf{T}\), the quadratic form is
If \(a\neq 0\), at least one of the first two squared terms is positive, so this expression is strictly positive. Thus the matrix is positive definite and has rank 2. As a numerical cross-check, its determinant is \(3(2)-1^2=5>0\); its diagonal entries are also positive. The determinant and diagonal check agrees with the quadratic-form proof for this \(2\times2\) example, but the quadratic-form argument directly establishes the general result.
Standardization Preserves the Rank
It is common to express the contrasts in standard-error units. Let \(S\) be the diagonal matrix whose \(i\)th diagonal entry is the standard deviation of \(D_i\), namely \(\sqrt{\sigma^2(1/n_i+1/n_0)}\). The correlation matrix of the standardized contrasts is \(R=S^{-1}\Sigma S^{-1}\). Since every diagonal entry of \(S\) is positive, \(S\) is invertible. For any vector \(b\),
Consequently, positive semidefiniteness and positive definiteness are preserved under this standardization. Multiplication on both sides by an invertible matrix also preserves rank, so \(R\) has rank \(m\) whenever \(\Sigma\) does. In particular, expressing the contrasts on a standardized scale does not create a linear dependence among them.
Worked Example: Rank of a Standardized Three-Contrast Matrix
Suppose, hypothetically, that \(\sigma^2=100\), \(n_0=40\), and the three treatment sample sizes are \(n_1=20\), \(n_2=30\), and \(n_3=60\). The contrast variances are \(7.5\), \(5.8333\), and \(4.1667\), and each covariance between distinct contrasts is \(2.5\). Hence
The correlation between the first and second contrasts, for example, is \(2.5/\sqrt{7.5(5.8333)}\approx 0.3780\). The diagonal standardization uses positive standard deviations, so it preserves the rank and definiteness of this covariance matrix. The quadratic-form result applies because all three treatment arms have finite positive sample sizes and \(\sigma^2>0\); therefore, this covariance matrix and its standardized correlation matrix both have rank 3.
As a numerical check, the determinant of this covariance matrix is approximately \(104.17\), which is positive. Dividing by the product of its three variances gives the determinant of its correlation matrix, approximately \(104.17/(7.5\times5.8333\times4.1667)=0.5714\), also positive. These calculations are consistent with full rank; the structural proof, rather than rounded numerical entries, establishes the exact result.
Diagnostics, Alternatives, and Limits
The quadratic-form definition is the most direct diagnostic: a proposed covariance matrix is positive semidefinite only if no coefficient vector produces a negative quadratic form. In practice, one may also inspect eigenvalues or principal minors. For a symmetric matrix, nonnegative eigenvalues characterize positive semidefiniteness; a negative eigenvalue proves that the matrix is invalid as a covariance matrix. Checking only the diagonal entries is insufficient: positive variances do not guarantee that the off-diagonal entries form a valid joint covariance structure.
A useful small-matrix check is to examine principal minors, meaning determinants of square submatrices formed using the same selected rows and columns. A negative principal minor rules out positive semidefiniteness. For larger matrices, eigenvalue calculations or the model-based quadratic-form proof are usually more informative than checking many minors individually. Numerical eigenvalues near zero should be interpreted with care: rounding, estimated covariance entries, or numerical tolerances can make a theoretically zero eigenvalue appear slightly positive or negative.
The alternatives serve different purposes. The model-based proof explains why the matrix must be valid and establishes its rank under stated assumptions. Eigenvalues and minors are useful checks on a particular numerical matrix, but do not explain whether its entries match the sampling design. Pairwise checks, such as confirming that each correlation lies between \(-1\) and \(1\), are weaker still: individually plausible correlations need not combine into a positive-semidefinite matrix.
Full rank is not guaranteed if the setup changes. If a contrast has no independent treatment-arm sampling variation, or if a list contains a contrast that is an exact linear combination of others, a nonzero coefficient vector can produce a zero-variance combination and the covariance matrix can be singular. Such a matrix remains positive semidefinite, but its rank is below its dimension. These cases do not contradict the full-rank result: they remove the positive treatment-specific variance or the distinctness of the contrasts required by that result.
Worked Example: Rejecting an Impossible Covariance Matrix
Suppose a hypothetical analysis proposes the matrix \(\begin{pmatrix}1&1.2\\1.2&1\end{pmatrix}\) as the covariance matrix of two quantities. Its diagonal entries are positive, but take \(a=(1,-1)^\mathsf{T}\). Then
This is negative, although the expression is supposed to be the variance of the difference between the two quantities. Therefore the proposed matrix is not positive semidefinite and cannot be a covariance matrix. Its determinant, \(1-1.2^2=-0.44\), provides a second warning. In a shared-control analysis, an invalid numerical matrix can indicate a calculation error, inconsistent entries, or an inappropriate model—not a legitimate consequence of sharing a control.
Common Pitfall
A frequent error is to infer that a shared control makes the contrast covariance matrix singular because every contrast contains the same control mean. Sharing does create covariance, but covariance is not the same as linear dependence. In the standard model, the treatment means contribute independent sampling variation to their respective contrasts. The quadratic form makes the distinction explicit: for nonzero \(a\), the term \(\sigma^2\sum_i a_i^2/n_i\) is strictly positive, even if \(\sum_i a_i=0\) and the shared-control term vanishes.
The correction is to distinguish the common source of covariance from the full random variation in each contrast. For finite, positive treatment sample sizes, and positive common variance, the \(m\) distinct treatment-control contrasts have rank \(m\). Positive semidefiniteness is the general requirement for a covariance matrix; positive definiteness and full rank follow here from the stronger independent treatment-arm sampling assumptions.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- For a shared-control contrast covariance matrix, what does positive semidefiniteness require of \(a^\mathsf{T}\Sigma a\) for every coefficient vector \(a\)?
- There are four distinct treatment-control contrasts, each treatment arm has finite positive sample size, and \(\sigma^2>0\). What is the rank of their covariance matrix, and why?
- Why does converting each contrast to standard-error units preserve the rank of the covariance matrix?
- A proposed covariance matrix has positive diagonal entries but has a negative eigenvalue. Can it be a valid covariance matrix? Explain.
Solutions. (1) It must satisfy \(a^\mathsf{T}\Sigma a\geq 0\) for every \(a\), because this quadratic form is the variance of the corresponding linear combination. (2) Its rank is 4. For every nonzero \(a\), the treatment-specific term \(\sigma^2\sum_i a_i^2/n_i\) is strictly positive, making the matrix positive definite and full rank. (3) Standardization multiplies on both sides by an invertible diagonal matrix of positive standard deviations; this preserves rank. (4) No. A negative eigenvalue means there is a direction with a negative quadratic form, implying a negative variance, which is impossible for a covariance matrix.