Wilks Lambda as a Determinant Ratio

Why a Determinant Ratio?

In MANOVA, a hypothesis concerns a collection of responses, so its effect is not summarized by a single scalar sum of squares. The hypothesis sum-of-squares-and-cross-products matrix \(H\) records variation attributed to the tested effect, while the error sum-of-squares-and-cross-products matrix \(E\) records residual variation. Wilks’ lambda compares the generalized volumes represented by these matrices. Its determinant form also makes the connection to the generalized roots exact.

The central identity is that the determinant ratio \(|E|/|E+H|\) equals a product of contributions from the MANOVA roots. A root near zero contributes a factor near one; a large root contributes a factor near zero. Thus the product decreases as any one root increases, with all other roots held fixed.

Definition and Conditions

Let there be \(p\) responses. Assume \(H\) is a symmetric positive-semidefinite \(p\times p\) hypothesis matrix and \(E\) is a symmetric positive-definite \(p\times p\) error matrix. Positive definiteness of \(E\) ensures that its determinant is positive and the generalized eigenvalue problem is well defined without restricting attention to a lower-dimensional subspace. Write \(|A|\) for the determinant of a square matrix \(A\).

To see why the roots have these properties, let \(E^{-1/2}\) denote the symmetric positive-definite inverse square root of \(E\), and define \(A=E^{-1/2}HE^{-1/2}\). This matrix is symmetric positive semidefinite: for every vector \(x\), \(x^\mathsf{T}Ax=(E^{-1/2}x)^\mathsf{T}H(E^{-1/2}x)\geq 0\). The generalized eigenvalues of \((H,E)\) are the ordinary eigenvalues of \(A\), so they are real and nonnegative. This symmetric representation is useful even when \(E^{-1}H\) itself is not symmetric.

From the Determinant Ratio to the Root Product

The determinant ratio and root product are two expressions for the same quantity. The proof uses the determinant’s multiplicative property and the eigenvalues of the symmetric matrix \(A\).

Proof. Factor \(E+H\) by a congruence involving \(E^{1/2}\):

Taking determinants and using \(|BCD|=|B||C||D|\) gives \(|E+H|=|E^{1/2}||I+A||E^{1/2}|=|E||I+A|\), since \(|E^{1/2}|^2=|E|\). Therefore \(\Lambda=1/|I+A|\). The eigenvalues of \(A\) are \(\lambda_1,\ldots,\lambda_p\), so the eigenvalues of \(I+A\) are \(1+\lambda_1,\ldots,1+\lambda_p\). The determinant is the product of the eigenvalues, giving

Every factor is positive and at most one, so \(0<\Lambda\leq1\). The product equals one if and only if each \(\lambda_i=0\), which for a positive-semidefinite \(A\) is equivalent to \(A=0\), and hence to \(H=0\). \(\square\)

If \(H\) has rank \(q<p\), then exactly \(p-q\) roots are zero. Those roots contribute factors of one, so only the nonzero roots change the product. The determinant expression still uses all \(p\) response dimensions.

Worked Determinant Calculations

How Increasing Roots Changes Wilks’ Lambda

The product formula gives a direct interpretation of a root’s contribution. Holding all other roots fixed, increasing \(\lambda_j\) increases the denominator factor \(1+\lambda_j\), and therefore decreases \(\Lambda\). More precisely, for nonnegative roots,

The log-scale derivative describes the proportional response: a small increase \(d\lambda_j\) changes \(\log\Lambda\) by approximately \(-d\lambda_j/(1+\lambda_j)\). Thus a fixed small increase has a smaller proportional effect when the root is already large. This does not make a large root unimportant; its factor \((1+\lambda_j)^{-1}\) is already small.

Monotonicity When the Hypothesis Matrix Increases

A matrix version of the same idea is useful when comparing two hypotheses with the same error matrix. For symmetric matrices, write \(H_2\succeq H_1\) when \(H_2-H_1\) is positive semidefinite. This means the increment adds nonnegative quadratic-form variation in every direction.

Proof. Put \(D=H_2-H_1\succeq0\), \(A=E^{-1/2}H_1E^{-1/2}\), and \(C=E^{-1/2}DE^{-1/2}\succeq0\). The determinant-ratio identity already proved gives \(\Lambda(H_1)=1/|I+A|\) and \(\Lambda(H_2)=1/|I+A+C|\).

Factor the second determinant as

Indeed, the matrix inside the determinant is obtained by multiplying \(I+A+C\) on both sides by \((I+A)^{-1/2}\), and the two outside determinant factors contribute \(1/|I+A|\). The matrix \(K=(I+A)^{-1/2}C(I+A)^{-1/2}\) is positive semidefinite. Its eigenvalues \(\kappa_i\) are nonnegative, so \(|I+K|=\prod_i(1+\kappa_i)\geq1\). It follows that \(|I+A+C|\geq|I+A|\), and taking reciprocals yields \(\Lambda(H_2)\leq\Lambda(H_1)\). If \(D\neq0\), then \(C\neq0\) because congruence by the invertible matrix \(E^{-1/2}\) cannot map a nonzero matrix to zero. Thus \(K\neq0\), at least one \(\kappa_i>0\), and \(|I+K|>1\); the inequality is strict. \(\square\)

Interpretation and Common Pitfall

Wilks’ lambda is a dimensionless determinant ratio comparing the error matrix with the error-plus-hypothesis matrix. Under the stated conditions it lies in \((0,1]\). A value near one means the product of root contributions is near one; a smaller value means that one or more roots have made the product smaller. This is a statement about the criterion’s algebraic response, not by itself a significance conclusion: interpreting a value as evidence in a test requires an appropriate null distribution or calibration.

• Pitfall: treating the roots as diagonal entries. In general, the roots are not the diagonal entries of \(H\), \(E\), or \(E^{-1}H\). They are generalized eigenvalues. The second worked example shows how to find them even when the displayed matrices are not diagonal.
• Pitfall: reversing the ratio. The convention here is \(|E|/|E+H|\), not its reciprocal. Since \(H\succeq0\), the denominator determinant is at least the numerator determinant, consistent with \(0<\Lambda\leq1\).
• Pitfall: assuming every root must be positive. If \(H\) is rank deficient, some roots are zero. Such roots contribute one to the product and do not change \(\Lambda\).