Can an eigenvalue be negative?

Yes. For a real eigenvector, it reverses orientation while scaling length by its absolute value.

Does every square real matrix have a real eigenvalue?

No. A real matrix can have only nonreal eigenvalues; over the complex numbers, its characteristic polynomial has roots.

What does an eigenvalue of zero imply?

The square matrix has a nonzero vector in its nullspace and is therefore singular.

AI Glossary / Mathematics and optimization / Intermediate

Eigenvalue

Definition

An eigenvalue is a number that describes how a square matrix scales a corresponding nonzero eigenvector. If Av = λv, then λ is the eigenvalue and v is an eigenvector. Along that vector, applying the matrix has the same effect as multiplying by λ. Eigenvalues can be real or complex, even when all entries of the matrix are real.

What it means

For a real eigenvalue and real eigenvector, the matrix preserves the vector's line through the origin. A positive eigenvalue keeps its orientation, a negative eigenvalue reverses it, and a zero eigenvalue sends it to zero. The magnitude tells you the length-scaling factor along that direction. This interpretation applies to the associated eigenvector, not automatically to every input vector. A matrix can stretch different directions differently, and a general input may combine several behaviors.

How it works

The equation Av = λv can be rewritten as (A − λI)v = 0, where I is the identity matrix. A nonzero solution exists exactly when A − λI is singular. For small examples, this leads to the characteristic equation det(A − λI) = 0. Once a candidate eigenvalue is found, solve the corresponding homogeneous system for its eigenvectors. Practical numerical software usually uses specialized matrix algorithms rather than explicitly expanding a high-degree determinant. Matrix structure matters: real symmetric matrices have real eigenvalues and admit an orthonormal eigenvector basis.

Three practical examples

Principal component analysis interprets eigenvalues of a covariance matrix as variances along the corresponding principal directions.
A linear discrete-time model can use eigenvalues to analyze how modes behave under repeated updates.
Optimization analysis can examine eigenvalues of a symmetric Hessian to study local curvature, subject to the conditions required for the particular conclusion.

Think of it this way

Think of a transformation with certain special tracks: an object placed exactly on one track stays on that track while being scaled. The eigenvalue is the scale factor for that track. The analogy works for real eigenvectors, including reversed orientation, but it does not fully describe complex eigenpairs or transformations without a complete set of independent eigenvectors. It should not be taken to mean every matrix has a set of real tracks spanning the whole space.

Worked example

Take A = [[4, 1], [1, 4]]. The characteristic equation is (4 − λ)² − 1 = 0, giving λ = 5 and λ = 3. Check v1 = [1, 1]: Av1 = [5, 5] = 5v1. Check v2 = [1, −1]: Av2 = [3, −3] = 3v2. Now decompose x = [2, 0] as v1 + v2. One application gives Ax = 5v1 + 3v2 = [8, 2]. Two applications give A²x = 25v1 + 9v2 = [34, 16]. The different eigenvalues explain how the two components grow at different rates. Their sum is 8, matching the trace 4 + 4; their product is 15, matching the determinant 16 − 1. If A were used as a valid covariance matrix for two equally scaled features, the leading principal direction would account for 5/(5 + 3) = 62.5% of total variance. That percentage follows from the covariance interpretation, not from eigenvalues having an inherent percentage meaning in every application.

Illustrative case study

In a fictional analysis, Laurel Research studies two centered measurement features with the illustrative covariance matrix above. Analysts report both principal directions and their variances, then check whether the feature units make the covariance comparison sensible. They do not call the larger eigenvalue a measure of business importance: a direction with substantial variation may still be irrelevant to the prediction target. The case illustrates interpretation of a matrix, not a claim that an invented dimensionality reduction improved accuracy.

Common misconceptions and mistakes

An eigenvalue is a scalar, not a vector. It may be zero even though an eigenvector may not be the zero vector. Real matrix entries do not guarantee real eigenvalues: planar rotations provide counterexamples. A repeated eigenvalue does not always provide enough independent eigenvectors to diagonalize a matrix. Eigenvalues are not generally identical to singular values, although special matrix classes connect them. Finally, a large eigenvalue has no universal good-or-bad meaning; it could indicate variance, amplification or curvature depending on what the matrix represents.

Business use

Before interpreting eigenvalues, identify the matrix and its units: covariance, transition, Hessian and adjacency matrices answer different questions. Use software suited to its structure, such as symmetry when present. Check an estimated eigenpair with the residual Av − λv and interpret the residual relative to relevant scales. In a stability analysis, distinguish discrete-time updates from continuous-time dynamics because their eigenvalue conditions differ. When conclusions depend on a small eigenvalue gap or a nearly zero value, examine numerical precision and sensitivity rather than turning a rounded number into an exact structural claim.

Check your understanding

For the worked A, does eigenvalue 5 mean every vector becomes five times longer?

Answer: No. The factor 5 applies along its associated eigendirection. The other demonstrated eigendirection is scaled by 3, and a general vector combines these effects.

Related terms

Eigenvector
Covariance matrix
Singular value
characteristic polynomial
diagonalization
spectral radius

Learn next

verify A²x directly by multiplying A by [8, 2]. Then replace A with a diagonal matrix having entries 2 and −1 and trace repeated applications to each coordinate direction. Finally examine a 90-degree rotation to understand why complex eigenvalues can be necessary.

Sources

Sources: TU Delft, Linear Algebra, Definitions and examples of eigenvalues and eigenvectors, https://interactivetextbooks.tudelft.nl/linear-algebra/Chapter6/EigenvaluesEigenvectors.html ; Jung and colleagues, The Aalto Dictionary of Machine Learning, eigenvalue, https://aaltodictionaryofml.github.io/ADictML_Main.pdf . The matrix, repeated-update calculation and fictional covariance analysis are original.

Previous: Eigenmoments Next: Eigenvector