Eigenvalue

What an eigenvalue reveals about a matrix

That direction is represented by an eigenvector, and together the pair can reveal long-run behaviour, dominant patterns and numerical stability that are difficult to see from individual matrix entries. This page explains the defining equation, shows how eigenvalues are calculated and interpreted, and connects them to state-transition models, covariance structures and other health-economic applications.

Eigenvalues are properties of square matrices. They summarise selected transformations produced by a matrix; they do not replace the underlying model, its assumptions or the interpretation of its rows and columns.

The equation that defines an eigenvalue

Suppose a square matrix (A) transforms a non-zero vector (v). If the result points in the same or exactly opposite direction as (v), differing only by a scalar multiplier, then (v) is an eigenvector of (A) and that multiplier is its eigenvalue. The defining equation is:

$$Av=\lambda v$$

where:

• (A) is an (n \times n) square matrix.
• (v) is a non-zero eigenvector.
• (\lambda) is the eigenvalue associated with (v).

The value of (\lambda) determines how the transformation changes the eigenvector:

How eigenvalues are calculated

The defining equation can be rearranged so that the eigenvalue problem becomes a condition on a determinant. A non-zero solution for (v) exists only when (A-\lambda I) is singular. This produces the characteristic equation:

$$\det(A-\lambda I)=0$$

The polynomial (\det(A-\lambda I)) is the characteristic polynomial, and its roots are the eigenvalues of (A). Once an eigenvalue is known, its associated eigenvectors are found by solving:

$$(A-\lambda I)v=0$$

For a (2\times2) matrix:

$$A= \begin{bmatrix} a & b \\ c & d \end{bmatrix}$$

the characteristic equation is:

$$\lambda^2-(a+d)\lambda+(ad-bc)=0$$

The coefficient (a+d) is the trace of (A), and (ad-bc) is its determinant. Therefore, counting algebraic multiplicity, the sum of the eigenvalues equals the trace and their product equals the determinant.

A worked two-state calculation

Consider a two-state transition matrix in which each row represents a current state and each column represents a next state. The calculation illustrates both how to obtain the eigenvalues and why one eigenvalue of a stochastic transition matrix is typically one. It also shows how the other eigenvalue governs the rate at which the effect of the initial state distribution diminishes.

Let:

$$P= \begin{bmatrix} 0.8 & 0.2 \\ 0.1 & 0.9 \end{bmatrix}$$

The characteristic equation is:

$$\det(P-\lambda I) = \begin{vmatrix} 0.8-\lambda & 0.2 \\ 0.1 & 0.9-\lambda \end{vmatrix} =0$$

Expanding the determinant gives:

$$(0.8-\lambda)(0.9-\lambda)-0.02=0$$ $$\lambda^2-1.7\lambda+0.70=0$$ $$(\lambda-1)(\lambda-0.7)=0$$

The eigenvalues are therefore (\lambda_1=1) and (\lambda_2=0.7). The eigenvalue one corresponds to the matrix's stable probability structure, while the magnitude (0.7) determines how quickly deviations from that structure contract under repeated multiplication.

What eigenvalues mean in a transition model

A cohort state-transition model repeatedly multiplies a state vector by a transition matrix. Eigenvalues help describe the behaviour of repeated powers of that matrix, including persistence, convergence and absorption. Their interpretation depends on the transition convention, the structure of the chain and whether the model is time-homogeneous.

For a row-stochastic transition matrix (P), each row sums to one and (\lambda=1) is an eigenvalue. Under suitable conditions, a stationary row vector (\pi) satisfies:

$$\pi P=\pi$$

with:

$$\sum_i \pi_i=1$$

The stationary distribution is therefore a left eigenvector associated with eigenvalue one. If a column-vector convention is used, the corresponding orientation changes, so analysts must not switch between left and right eigenvectors without also changing the matrix convention.

In a finite irreducible and aperiodic Markov chain, the other eigenvalues have magnitude below one and their contributions decay as the matrix is repeatedly applied. The second-largest eigenvalue magnitude often indicates the speed of convergence: a value close to one implies slow convergence, while a smaller magnitude implies faster convergence.

Absorbing states change the long-run interpretation

Health-economic transition models commonly include an absorbing state such as death. Once the cohort enters that state, it cannot leave, so the model's long-run distribution may be driven by absorption rather than by a clinically meaningful equilibrium among living states. An eigenvalue of one is therefore expected, but it must not automatically be interpreted as evidence of a useful steady-state population.

For an absorbing state (j):

$$p_{jj}=1$$

and:

$$p_{jk}=0 \quad \text{for } k\ne j$$

A model with more than one closed communicating class may have eigenvalue one with multiplicity greater than one. In that situation, the limiting distribution can depend on the initial state and on which closed class is reached.

The dominant eigenvalue and spectral radius

The spectral radius is the largest absolute value among a matrix's eigenvalues. It provides a compact indication of how repeated multiplication may amplify, preserve or reduce components of a vector. The spectral radius is defined as:

$$\rho(A)=\max_i |\lambda_i|$$

If (\rho(A)<1), then (A^t) approaches the zero matrix as (t) increases. If (\rho(A)>1), repeated multiplication can cause at least one component to grow in magnitude, while (\rho(A)=1) requires closer examination of multiplicity and matrix structure before long-run behaviour can be concluded.

For a stochastic transition matrix, the spectral radius is one. This result follows from the probability-preserving structure of the matrix, but it does not by itself establish irreducibility, convergence to a unique stationary distribution or clinical plausibility.

How eigenvalues support uncertainty analysis

Eigenvalues are also used to examine variance-covariance and correlation matrices. Because a valid covariance matrix must be positive semidefinite, all of its eigenvalues should be non-negative apart from negligible deviations caused by numerical rounding. A materially negative eigenvalue can indicate an invalid covariance specification that will cause problems in correlated parameter sampling.

For a covariance matrix (\Sigma):

$$x^{\mathsf T}\Sigma x\ge 0 \quad \text{for every vector } x$$

This condition is equivalent to all eigenvalues of (\Sigma) being non-negative. When a probabilistic sensitivity analysis samples correlated parameters, checking the eigenvalues can therefore detect incompatible correlations, inconsistent rounding or an incorrectly assembled matrix before simulation begins.

Small positive eigenvalues may also signal near-linear dependence. The matrix can remain mathematically valid while being numerically fragile, so apparently minor changes in inputs may produce large changes in inverses, decompositions or sampled values.

How principal component analysis uses eigenvalues

Principal component analysis reduces a set of correlated variables to a smaller number of uncorrelated components. Each component corresponds to an eigenvector of a covariance or correlation matrix, and its eigenvalue measures the variance explained along that direction. This can help summarise high-dimensional health, utilisation or preference data, although the resulting components still require substantive interpretation.

If the ordered eigenvalues are (\lambda_1\ge\lambda_2\ge\cdots\ge\lambda_n\ge0), the proportion of total variance explained by component (k) is:

$$\frac{\lambda_k}{\sum_{i=1}^{n}\lambda_i}$$

Retaining components solely because their eigenvalues exceed an arbitrary threshold can be misleading. The number retained should also reflect the analysis purpose, data-generating context, interpretability and validation performance.

Repeated and complex eigenvalues need careful handling

An eigenvalue may occur more than once as a root of the characteristic polynomial. Its algebraic multiplicity counts how often it appears as a root, while its geometric multiplicity is the number of linearly independent eigenvectors associated with it. When a matrix does not have enough independent eigenvectors, it cannot be diagonalised in the usual way.

A real matrix can also have complex eigenvalues, which occur in conjugate pairs. Complex values are not automatically errors; they can reflect rotational or oscillatory behaviour. In a health-economic model, however, analysts should confirm that the matrix being analysed is the intended object and that any resulting interpretation remains meaningful for the clinical process.

Why software results can differ slightly

Analytical formulas are useful for small matrices, but applied models normally obtain eigenvalues using numerical algorithms. Floating-point arithmetic, matrix scaling and the choice of algorithm can produce tiny imaginary parts or very small negative values that are numerical artefacts rather than substantive features. The size of the discrepancy should be judged relative to the matrix scale and the purpose of the calculation.

Directly solving a high-degree characteristic polynomial is generally less stable than using established numerical eigensolvers. Analysts should preserve adequate precision, avoid premature rounding and compare suspicious results across a second implementation or decomposition.

Common interpretation errors

Eigenvalues compress important information, but they do not explain a matrix without its eigenvectors, labels and modelling context. A technically correct eigenvalue can be misinterpreted if the analyst ignores orientation, multiplicity or the assumptions behind the matrix. The following errors are especially important in health-economic applications.

• Treating every eigenvalue of one as a unique equilibrium ignores reducibility, absorbing states and multiplicity.
• Confusing left and right eigenvectors can attach the stationary interpretation to the wrong vector orientation.
• Using only the largest eigenvalue can hide slow convergence, oscillation or multiple closed classes represented by other eigenvalues.
• Interpreting a negative covariance eigenvalue without checking scale can mistake harmless floating-point noise for a material validity problem, or dismiss a genuinely invalid covariance matrix as rounding.
• Assuming diagonalisation is always possible ignores matrices with insufficient independent eigenvectors.
• Applying time-homogeneous conclusions to changing matrices is inappropriate when transition probabilities vary by cycle, age or time since treatment.
• Equating long-run mathematical behaviour with decision relevance ignores finite horizons, discounting, competing risks and clinical plausibility.

A practical validation sequence

An eigenvalue analysis should begin by confirming that the matrix itself represents the intended quantities. Validation should then test mathematical identities, numerical stability and the substantive meaning of the result. This sequence makes the calculation auditable without treating the software output as self-validating.

1. Confirm that the matrix is square and verify every row label, column label and ordering convention.
2. Check the matrix-specific constraints, including stochastic row sums or covariance symmetry and positive semidefiniteness.
3. Calculate a small case independently and verify the characteristic equation, trace and determinant relationships.
4. Substitute each eigenpair into (Av=\lambda v) and examine the residual (Av-\lambda v).
5. Inspect multiplicity and orientation before interpreting stationary distributions or long-run behaviour.
6. Compare eigenvalue magnitudes to assess persistence, convergence or numerical sensitivity.
7. Repeat suspicious calculations at higher precision or with a second reliable numerical method.
8. Relate the result back to the model horizon and decision problem rather than reporting a purely mathematical conclusion.

Eigenvalues are most useful when they reveal behaviour that can be traced back to an explicit model structure. They should support, not replace, validation of transition probabilities, covariance assumptions, clinical pathways and economic outputs.