Tutorials › CAS Exam 9 - Risk Management for Actuaries › Checking Covariance Matrix Validity

Home  /  Tutorials  /  CAS Exam 9 / Aggregating Risks and Diversification

CAS Exam 9 · Aggregating Risks and Diversification · Tutorial 474 of 995

Checking Covariance Matrix Validity

You will check covariance matrices for internal consistency using eigenvalues, determinants, and principal minors.

Advanced 11 min read

What You'll Learn

  • Define covariance and correlation matrices and the positive semidefinite condition.
  • Explain why every linear combination of risks must have nonnegative variance.
  • Use eigenvalues to detect impossible dependence inputs.
  • Apply principal-minor checks, including the distinction between positive semidefinite and positive definite.
  • Recognize why valid pairwise correlations alone do not guarantee a valid matrix.
  • Interpret singular matrices and small negative eigenvalues in rounded or estimated inputs.

Why Covariance Matrix Validity Matters

When risks are aggregated, the covariance matrix summarizes their variances and pairwise covariances. Those entries cannot be chosen independently without restriction: together, they must describe a possible joint distribution. An inconsistent matrix can imply that some portfolio of risks has negative variance, which is impossible regardless of the marginal models or the dependence method used to produce the inputs.

The central validity condition is positive semidefiniteness. It is a mathematical requirement on the matrix, not an assumption that the risks are normally distributed. It applies whenever the risks have finite second moments.

Definition: For random variables \(X_1,\ldots,X_n\) with finite variances, their covariance matrix is \(\Sigma\), with entries \(\Sigma_{ij}=\operatorname{Cov}(X_i,X_j)\). It is symmetric, and \(\Sigma_{ii}=\operatorname{Var}(X_i)\). The matrix is positive semidefinite if \(a^\mathsf{T}\Sigma a\geq 0\) for every real vector \(a\). It is positive definite if \(a^\mathsf{T}\Sigma a>0\) for every nonzero \(a\).

A correlation matrix is the corresponding standardized matrix when every risk has positive standard deviation: its diagonal entries are 1, and its \(i,j\) entry is \(\operatorname{Corr}(X_i,X_j)\). A covariance matrix must be positive semidefinite, as must a correlation matrix. The correlation matrix can be obtained from the covariance matrix by dividing each entry by the product of the relevant standard deviations.

Why the Condition Must Hold

For any chosen weights \(a_1,\ldots,a_n\), consider the linear combination \(Z=\sum_i a_iX_i\). Its variance cannot be negative. Expanding that variance gives exactly the quadratic form used in the definition.

$$ \begin{aligned} \operatorname{Var}(Z) &=\operatorname{Var}\left(\sum_{i=1}^n a_iX_i\right)\\ &=\sum_{i=1}^n\sum_{j=1}^n a_i a_j\operatorname{Cov}(X_i,X_j)\\ &=a^\mathsf{T}\Sigma a\geq 0. \end{aligned} $$

The first equality follows by writing the variance as the expected squared deviation from the mean; distributing the product of the two sums produces one covariance term for each pair \(i,j\). This also explains why an indefinite matrix is impossible: a vector \(a\) with \(a^\mathsf{T}\Sigma a<0\) would specify a linear combination with negative variance.

For a real symmetric matrix, the spectral theorem gives real eigenvalues and an orthonormal set of eigenvectors. The matrix is positive semidefinite exactly when every eigenvalue is nonnegative. In fact, if \(u\) is an eigenvector with eigenvalue \(\lambda\), then \(u^\mathsf{T}\Sigma u=\lambda u^\mathsf{T}u\). A negative eigenvalue therefore directly supplies a direction with negative quadratic form.

Worked Example: Detecting an Impossible Correlation Matrix by Eigenvalues

Worked Example: Detecting an Impossible Correlation Matrix by Eigenvalues

A model supplies three standardized risk changes with correlations \(\rho_{12}=0.9\), \(\rho_{13}=0.9\), and \(\rho_{23}=-0.9\). Each pairwise value lies between \(-1\) and \(1\). Determine whether the complete set of inputs is valid, using eigenvalues, and interpret any failure.

Method and assumptions. Treat the stated values as entries of a proposed correlation matrix. Any valid correlation matrix must be symmetric and positive semidefinite. The diagonal entries are 1 because each variable is standardized.

$$ R= \begin{pmatrix} 1&0.9&0.9\\ 0.9&1&-0.9\\ 0.9&-0.9&1 \end{pmatrix} $$

The vector \(u=(1,-1,-1)^\mathsf{T}\) is an eigenvector:

$$ Ru= \begin{pmatrix} 1-0.9-0.9\\ 0.9-1+0.9\\ 0.9+0.9-1 \end{pmatrix} = \begin{pmatrix} -0.8\\ 0.8\\ 0.8 \end{pmatrix} =-0.8u. $$

Thus one eigenvalue is \(-0.8\). The other two eigenvalues are \(1.9\) and \(1.9\): for example, vectors \((1,1,0)^\mathsf{T}\) and \((1,-1,2)^\mathsf{T}\) are eigenvectors with eigenvalue \(1.9\). The eigenvalues sum to the trace, \(3\), as a check: \(-0.8+1.9+1.9=3\).

Conclusion. Since one eigenvalue is negative, \(R\) is not positive semidefinite and the three correlations cannot all belong to the same collection of risks. The contradiction is also visible in the variance calculation: for these standardized risks, the combination \(X_1-X_2-X_3\) would have variance \(u^\mathsf{T}Ru=-0.8(3)=-2.4\), which is impossible.

Common mistake avoided. Checking that each individual correlation lies in \([-1,1]\) is necessary, but not sufficient. It does not test whether all three pairwise relationships can hold simultaneously.

Principal-Minor Checks

A principal minor is the determinant of a submatrix formed by keeping the same set of row and column indices. For a symmetric matrix, all principal minors must be nonnegative for the matrix to be positive semidefinite. For a positive definite matrix, it is sufficient to check that all leading principal minors are positive; that familiar shortcut must not be confused with the positive semidefinite test. In particular, checking only the diagonal and the leading principal minors can miss a failure.

For a \(3\times3\) correlation matrix with off-diagonal entries \(r_{12},r_{13},r_{23}\), its determinant is:

$$ \det(R)=1+2r_{12}r_{13}r_{23} -r_{12}^2-r_{13}^2-r_{23}^2. $$

The \(1\times1\) principal minors are the diagonal entries. Each \(2\times2\) principal minor for a correlation matrix has the form \(1-r_{ij}^2\). These checks test the individual variances and pairs; the \(3\times3\) determinant adds a joint constraint that cannot be reduced to pairwise bounds alone.

Worked Example: Applying Principal-Minor Checks to Covariance Inputs

Three hypothetical risk estimates have standard deviations 2, 3, and 4 (in the same units). The supplied correlations are 0.6 between risks 1 and 2, 0.6 between risks 1 and 3, and \(-0.6\) between risks 2 and 3. Check the resulting covariance matrix using principal minors.

Method and assumptions. Convert correlations to covariances using \(\operatorname{Cov}(X_i,X_j)=\rho_{ij}\sigma_i\sigma_j\). A valid covariance matrix must have nonnegative principal minors. The diagonal entries are the variances \(2^2=4\), \(3^2=9\), and \(4^2=16\).

The off-diagonal covariances are \(0.6(2)(3)=3.6\), \(0.6(2)(4)=4.8\), and \(-0.6(3)(4)=-7.2\). Therefore,

$$ \Sigma= \begin{pmatrix} 4&3.6&4.8\\ 3.6&9&-7.2\\ 4.8&-7.2&16 \end{pmatrix}. $$

The \(1\times1\) principal minors are \(4\), \(9\), and \(16\), all positive. The three \(2\times2\) principal minors are:

$$ \begin{aligned} 4(9)-(3.6)^2&=23.04,\\ 4(16)-(4.8)^2&=40.96,\\ 9(16)-(-7.2)^2&=92.16. \end{aligned} $$

All are positive, so every pairwise check passes. But the determinant of the full matrix is:

$$ \begin{aligned} \det(\Sigma) &=4\{9(16)-(-7.2)^2\} -3.6\{3.6(16)-(-7.2)(4.8)\}\\ &\quad+4.8\{3.6(-7.2)-9(4.8)\}\\ &=4(92.16)-3.6(92.16)+4.8(-69.12)\\ &=368.64-331.776-331.776\\ &=-294.912. \end{aligned} $$

Conclusion. The negative determinant means the matrix is not positive semidefinite. The pairwise covariance inputs are individually plausible but collectively inconsistent. The calculation is consistent with the standardized check: the correlation matrix has determinant \(1+2(0.6)(0.6)(-0.6)-3(0.6^2)=-0.512\), and scaling by standard deviations multiplies its determinant by \((2\cdot3\cdot4)^2=576\), giving \(-0.512(576)=-294.912\).

Common mistake avoided. Passing all \(1\times1\) and \(2\times2\) checks does not establish validity for three or more risks. The full determinant is a principal minor too and must also be nonnegative.

A Valid Matrix Can Be Singular

Positive semidefinite does not mean positive definite. A zero eigenvalue or zero principal minor is permitted. It can arise when a nonzero linear combination of risks has exactly zero variance, such as when one risk is an exact linear function of others. Such a matrix is singular, but it is still a valid covariance matrix. Requiring every eigenvalue to be strictly positive would incorrectly reject this case.

Worked Example: Verifying a Valid, Nonsingular Covariance Matrix

A hypothetical three-risk portfolio has standard deviations 10, 20, and 30. The correlations between risks 1 and 2 and between risks 2 and 3 are each 0.5; the correlation between risks 1 and 3 is 0. Check validity using principal minors and confirm that the matrix is nonsingular.

Method and assumptions. Convert the stated correlations to covariances. Then test all principal minors; strict positivity will establish positive definiteness in this case, and a positive full determinant will show the matrix is nonsingular.

The variances are \(100\), \(400\), and \(900\). The nonzero covariances are \(0.5(10)(20)=100\) and \(0.5(20)(30)=300\); the remaining covariance is zero. Hence:

$$ \Sigma= \begin{pmatrix} 100&100&0\\ 100&400&300\\ 0&300&900 \end{pmatrix}. $$

The \(1\times1\) principal minors are \(100\), \(400\), and \(900\). The \(2\times2\) principal minors are:

$$ \begin{aligned} 100(400)-100^2&=30{,}000,\\ 100(900)-0^2&=90{,}000,\\ 400(900)-300^2&=270{,}000. \end{aligned} $$

Expanding the full determinant along the first row gives:

$$ \det(\Sigma) =100\{400(900)-300^2\} -100\{100(900)-0(300)\} =27{,}000{,}000-9{,}000{,}000 =18{,}000{,}000. $$

Conclusion. Every principal minor is strictly positive, so the matrix is positive definite and therefore positive semidefinite. Its determinant is positive, so it is nonsingular. The corresponding correlation matrix has determinant \(0.5\); multiplying by \((10\cdot20\cdot30)^2=36{,}000{,}000\) again gives \(18{,}000{,}000\), checking the calculation.

Common mistake avoided. A positive diagonal alone would only show that each standalone variance is positive. The pairwise minors and the full determinant are needed to check whether the covariances are mutually compatible.

Practical Checks and Interpretation

For an actuarial aggregation model, validity checks should be applied to the matrix that will actually be used, after risks have been aligned to a common basis. A covariance matrix should be symmetric, have nonnegative diagonal entries, and be positive semidefinite. If a correlation matrix is supplied, check its unit diagonal and symmetric entries as well. Symmetry and acceptable individual correlations are necessary checks, not substitutes for a full matrix test.

For larger matrices, computing all principal minors becomes cumbersome. Eigenvalue software provides a practical test: a symmetric matrix is positive semidefinite if and only if all its eigenvalues are nonnegative. Principal minors remain especially useful for small matrices and for explaining exactly which condition fails. The equivalent quadratic-form check can also diagnose the failure: a negative value \(a^\mathsf{T}\Sigma a\) identifies a linear combination with an impossible implied variance.

Numerical inputs are often rounded or estimated. A tiny negative eigenvalue close to zero can result from rounding or numerical tolerance, so it should be assessed against the precision and scale of the data. It is not a reason to silently accept a materially negative eigenvalue. Any adjustment to make a matrix positive semidefinite changes the dependence inputs and should be documented and validated against the modeling purpose. A zero eigenvalue, by contrast, is mathematically allowed and may reflect exact dependence.

Key takeaway: A covariance matrix is valid only if every linear combination of the modeled risks has nonnegative variance. Test the whole matrix, not just its entries or its pairs: eigenvalues must be nonnegative, and all principal minors must be nonnegative.

Check Your Understanding

For each proposed input, decide whether it is positive semidefinite. Use the stated check and show enough work to support your conclusion.

  1. A two-risk correlation matrix has off-diagonal entry \(0.7\). What is its determinant, and is it valid? (Answer: determinant \(0.51\); valid.)
  2. A three-risk correlation matrix has off-diagonal entries \(r_{12}=0.5\), \(r_{13}=0.5\), and \(r_{23}=-0.5\). Check its determinant and determine whether it is positive semidefinite. (Answer: determinant \(0\); positive semidefinite and singular.)
  3. A covariance matrix is \(\begin{pmatrix}9&6\\6&4\end{pmatrix}\). Check its principal minors. What does the result say about validity? (Answer: principal minors are \(9\), \(4\), and \(0\); valid and singular.)
  4. A three-risk correlation matrix has off-diagonal entries \(0.4\), \(0.4\), and \(-0.4\). Use the \(3\times3\) determinant formula and the pairwise minors to assess validity. (Answer: pairwise minors are \(0.84\); determinant is \(0.392\); valid.)
  5. A three-risk correlation matrix has off-diagonal entries \(r_{12}=0.6\), \(r_{13}=0.6\), and \(r_{23}=-0.9\). Each pairwise minor is nonnegative. Does that establish validity? Check the full determinant. (Answer: no; determinant is \(-1.178\), so the matrix is invalid.)

Previous

Next

Back to Tutorials
Clinical Biostats