Can an eigenvector contain negative entries?

Yes. Its entries depend on the direction and coordinate system; negative values are not an error by themselves.

Can two eigenvectors have the same eigenvalue?

Yes. Every nonzero multiple already does, and some eigenspaces contain several independent directions.

Does every matrix have a complete eigenvector basis?

No. Some square matrices are not diagonalizable, and real eigenvectors alone may be insufficient when eigenvalues are complex.

AI Glossary / Mathematics and optimization / Intermediate

Eigenvector

Definition

An eigenvector is a nonzero vector that a square matrix transforms into a scalar multiple of itself. In the equation Av = λv, v is the eigenvector and λ is its eigenvalue. For a real eigenpair, the transformation keeps the vector on the same line through the origin, although it can stretch it, reverse its orientation or collapse it to zero.

What it means

The eigenvector identifies a special direction or mode of the transformation. Its length is usually a convention: if v is an eigenvector, every nonzero scalar multiple of v is also an eigenvector for the same eigenvalue. This is why software often reports unit-length eigenvectors and why two valid results can differ by a minus sign. The zero vector is excluded because A0 = λ0 for every λ, which would make it incapable of identifying a special scaling behavior.

How it works

After finding an eigenvalue λ, solve (A − λI)v = 0. All solutions form a nullspace; its nonzero members are the eigenvectors associated with λ. Adding the zero vector gives the full eigenspace. Different eigenvalues have linearly independent eigenvectors, but repeated eigenvalues require care: their eigenspaces can have different dimensions. A matrix is diagonalizable only when it has enough independent eigenvectors to form a basis. For real symmetric matrices, an orthonormal eigenvector basis is available, which makes many geometric interpretations especially convenient.

Three practical examples

Principal component analysis uses covariance eigenvectors to define directions for projecting centered observations.
A Markov-chain calculation can identify an invariant distribution using an eigenvector for eigenvalue 1, with the correct row or column convention and probability normalization.
A graph-analysis method can use selected eigenvectors of a graph matrix to build an embedding, with meaning determined by the matrix and algorithm rather than by the word eigenvector alone.

Think of it this way

Think of an arrow drawn on a flexible sheet. Most arrows change their direction when the sheet is transformed, but a special arrow remains on its original line and changes only by a scale factor. That arrow illustrates a real eigenvector. The analogy is limited to linear transformations about the origin and does not fully capture complex eigenvectors, defective matrices or application-specific meanings such as stationary probabilities.

Worked example

Let A = [[3, 2], [2, 3]]. For v = [1, 1], multiplication gives Av = [5, 5] = 5v. For w = [1, −1], Aw = [1, −1] = w. Thus v and w are eigenvectors with eigenvalues 5 and 1. Their unit-length versions are u1 = [1, 1]/√2 and u2 = [1, −1]/√2. They are perpendicular because their dot product is zero. Consider x = [3, 1]. Its coordinates along these unit vectors are u1ᵀx = 4/√2 = 2√2 and u2ᵀx = 2/√2 = √2. Hence x = 2√2u1 + √2u2. Keeping only the first component gives the projection [2, 2], leaving residual [1, −1] with squared length 2. If software returns −u1 instead, the projection coefficient changes to −2√2, but their product still gives [2, 2]. This explains why an eigenvector sign flip need not change a reconstruction. The projection has a principal-component interpretation if A is the relevant covariance matrix and x is expressed in the corresponding centered feature coordinates.

Illustrative case study

In a fictional reporting project, Cedar Surveys uses principal components to summarize two centered numeric features. After a software update, the first reported eigenvector has the opposite sign. Analysts compare the direction, scores and reconstruction together and confirm that the representation is equivalent after the matching score sign change. They retain a display convention for consistent charts. The case illustrates a harmless representation ambiguity; it does not assume every change in an eigenvector is harmless or that the components explain causation.

Common misconceptions and mistakes

An eigenvector is not generally unique, even after choosing an eigenvalue. Unit normalization fixes length but still permits a sign change in a one-dimensional real eigenspace. Repeated eigenvalues can permit rotations among basis vectors within the same eigenspace. Close eigenvalues can also make individual estimated directions sensitive to small data changes. Eigenvectors of a general matrix need not be orthogonal. A row-stochastic transition matrix's stationary distribution is a left eigenvector, so blindly using a right-eigenvector routine can answer the wrong question.

Business use

When presenting eigenvectors, state the matrix, normalization, ordering and coordinate meanings. For a data projection, preserve the centering and scaling used to build that matrix. Compare subspaces when repeated or nearly repeated eigenvalues make individual vector comparisons unstable. Validate a computed eigenpair using its residual and inspect whether the relevant eigenvalue is well separated. In business interpretation, examine the coefficients and the underlying variables before assigning a label to a component. A mathematical direction of variation is not automatically an actionable or causal factor.

Check your understanding

If u1 is replaced by −u1 in the worked projection, must the reconstructed vector change?

Answer: No. Its projection coefficient also changes sign, so multiplying the new coefficient by the new unit vector gives the same projected vector.

Related terms

Eigenvalue
Nullspace
Principal component analysis
eigenspace
diagonalization
orthonormal basis

Learn next

verify both matrix - vector products and reconstruct x from its two projection coefficients. Then compare the identity matrix, where every nonzero vector is an eigenvector, with a shear matrix that lacks a full eigenvector basis. Explain why eigenvector uniqueness depends on the matrix.

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, eigenvector, https://aaltodictionaryofml.github.io/ADictML_Main.pdf . The projection calculation and fictional sign-convention review are original educational examples.

Previous: Eigenvalue Next: Einsum notation