Eigenvalues and eigenvectors

Find eigenmodes, principal axes, and characteristic states. Theory, formulas, worked examples and interactive graphs.

What it is

An eigenvector of a square matrix A is a nonzero vector v that, when transformed by A, does not change direction: it is only rescaled. That scale factor is the associated eigenvalue λ, and both satisfy the equation Av = λv. Finding eigenvalues and eigenvectors means searching for the privileged directions of a linear transformation, those where the effect of A reduces to a simple stretch, compression, or sign flip.

This tool turns an apparently complicated multidimensional transformation into a set of independent one-dimensional behaviors, one per eigenvector. Instead of tracking how coordinates mix under A, it is enough to track how each eigenvector is scaled.

Why it appears in physics

It appears because many physical systems are described by matrices acting on state, position, or velocity vectors, and the natural behavior of the system is determined by its invariant directions. In coupled oscillations, the eigenvectors are the normal modes, that is, the coordinate combinations that oscillate in sync at a single natural frequency related to the eigenvalue.

It also appears in rigid-body mechanics, where the eigenvectors of the inertia tensor are the principal axes of rotation, and in quantum mechanics, where the eigenvalues of an operator represent the possible values of a physical measurement and the eigenvectors are the states associated with that measurement. In the stability of dynamical systems, the sign and magnitude of the eigenvalues determine whether a perturbation grows, decays, or oscillates.

Visual or geometric idea

The visual idea is this: apply a linear transformation to a set of vectors arranged as a circle. Most change both direction and length, but special directions exist where the transformed vector lies on the same line as the original, only longer, shorter, or reversed. Those lines are the eigen-directions, and the factor by which the length changes is the eigenvalue.

When a matrix has several independent eigenvectors, those directions act as a natural set of axes for the system: if the problem is described using those axes instead of the original ones, the transformation becomes diagonal and each coordinate evolves separately.

Minimal symbolic language

Symbolically, one seeks λ and v ≠ 0 such that Av = λv, which is equivalent to (A − λI)v = 0. For a nontrivial solution to exist, the determinant of (A − λI) must be zero: det(A − λI) = 0. This equation is the characteristic polynomial of A, and its roots are the eigenvalues.

Once an eigenvalue λ is known, the corresponding eigenvector is obtained by solving the homogeneous system (A − λI)v = 0, that is, by finding the null space of the matrix A − λI. Eigenvectors are defined up to a scale factor, so they are usually normalized to unit length.

What the result means

The result means that, if the system lies exactly along an eigenvector direction, its evolution under A is as simple as possible: a direct multiplication by λ. An eigenvalue larger than one in magnitude indicates growth along that direction, one smaller than one indicates attenuation, a negative one indicates a sign flip on each application, and a complex eigenvalue signals oscillatory or rotational behavior.

In physical systems, the full set of eigenvalues describes the global qualitative behavior: natural vibration frequencies, the stability of an equilibrium, or the possible values of a measurable quantum quantity. If you cannot say what each eigenvalue and eigenvector represents physically, the calculation is not yet interpreted.

When not to use it

Do not use this tool when the matrix is not square, since the equation Av = λv has no dimensional meaning in that case. Also do not assume that every square matrix has as many independent eigenvectors as its size: some matrices are defective and cannot be diagonalized, even though they always have eigenvalues.

Do not force real eigenvalues either when the physical problem explicitly requires oscillatory behavior and the matrix produces complex eigenvalues: in that case, the real part describes growth or decay and the imaginary part describes the oscillation frequency, and both must be interpreted together.

Operational summary

Set up (A − λI)v = 0, solve the characteristic polynomial to obtain the eigenvalues, and for each eigenvalue solve the homogeneous system to obtain its associated eigenvector. Verify each pair by substituting directly into Av = λv. Interpret each eigenvalue as a physical scale factor and each eigenvector as an invariant direction of the system.

Procedure

Before calculating

First check that A is square, since only square matrices have eigenvalues and eigenvectors in the standard sense. Identify the size of the matrix and anticipate the maximum number of eigenvalues it can have: an n×n matrix has at most n eigenvalues, counting multiplicity and allowing complex values.

Also decide whether the physical context requires real eigenvalues (for example, measurable frequencies or energies) or admits complex ones (for example, systems with damping or rotation).

Step 1: identify the mathematical object

The central mathematical object is the characteristic polynomial p(λ) = det(A − λI). Before calculating anything, recognize that you are looking for the roots of that polynomial, not a direct numerical solution of an ordinary linear system.

If the matrix is symmetric, anticipate that all its eigenvalues will be real and that an orthogonal eigenvector basis exists; if it is not, be prepared for the possibility of complex conjugate eigenvalues.

Step 2: choose the operation

For 2×2 matrices, use the shortcut λ^2 − (trace of A) λ + det(A) = 0, which avoids expanding the determinant by hand. For 3×3 matrices or larger, expand the determinant of A − λI by cofactors, carefully grouping the terms in λ.

Once the characteristic polynomial is obtained, the next operation is to solve it for λ: factoring, the quadratic formula, or searching for rational roots, depending on the degree.

Step 3: preserve units and signs

Subtract λ only from the entries on the main diagonal of A: the matrix A − λI is formed by subtracting the identity matrix multiplied by λ, not a loose scalar from a single entry. A sign error at this step invalidates the whole downstream calculation.

If the entries of A carry physical units (for example, an inertia tensor or a coupling matrix with elastic constants), the eigenvalues inherit those units, and it is worth checking that the final result makes dimensional sense.

Step 4: calculate

For each root λ of the characteristic polynomial, substitute into (A − λI)v = 0 and solve the resulting homogeneous system. The matrix A − λI is singular for that λ by construction, so the system always has infinitely many proportional solutions.

Choose a free parameter, express the remaining components in terms of it, and write the eigenvector with the simplest possible value for that parameter, usually 1. Normalize to unit length only if the problem explicitly requires it.

Step 5: interpret the result

Substitute the obtained pair (λ, v) directly into Av = λv and check that both sides match component by component. This verification catches sign or algebra errors before the result is accepted as final.

Then translate each eigenvalue and eigenvector into the physical language of the problem: natural frequency, principal axis, eigenstate, or growth rate, as appropriate for the context.

Self-check checklist

Check that the matrix is square, that the characteristic polynomial has the correct degree, that each eigenvector satisfies Av = λv, and that you can state in words what each obtained eigenvalue represents physically. If any eigenvalue is complex, check whether the physical problem allows it or whether it signals oscillation rather than pure growth.

Guided example

Physical situation

Two equal masses connected by springs oscillate together. Their dynamics are described by a matrix A = [[2, −1], [−1, 2]] (in units where k/m = 1), which relates the restoring acceleration to the positions of both masses. The goal is to find the normal modes of oscillation, that is, the coordinate combinations that oscillate in sync.

Mathematical translation

Finding the normal modes is equivalent to finding the eigenvalues and eigenvectors of A. Each eigenvalue will be proportional to the square of a natural frequency, and each eigenvector will describe how the two masses move relative to each other in that mode.

Step-by-step calculation

The characteristic polynomial is λ^2 − (trace) λ + det(A) = λ^2 − 4λ + 3, whose roots are λ = 1 and λ = 3. For λ = 1, the system (A − I)v = 0 gives v1 − v2 = 0, that is, v = (1, 1): both masses move together, in phase. For λ = 3, the system (A − 3I)v = 0 gives v1 + v2 = 0, that is, v = (1, −1): the masses move in opposite directions, out of phase.

Reading the result

The smaller eigenvalue, λ = 1, corresponds to the slower mode, where both masses oscillate together without compressing the middle spring. The larger eigenvalue, λ = 3, corresponds to the faster mode, where the middle spring is compressed and stretched to its maximum, raising the natural frequency. The frequency of each mode is proportional to the square root of its eigenvalue.

Common error avoided

This avoids treating the coupled system coordinate by coordinate, which would produce mixed differential equations impossible to solve separately. By switching to the eigenvector basis, each mode decouples and is solved as an independent oscillator.

Inverse example

Given result

A system has an eigenvalue λ = 0 with eigenvector v = (1, 1, 1).

What physical question it can answer

A zero eigenvalue indicates a direction where the transformation collapses: any vector proportional to (1, 1, 1) is sent to the zero vector. Physically, this usually signals a mode with no restoring force, a direction of neutral equilibrium, or a conserved quantity in a stability problem, where a perturbation along that specific direction generates no returning force and no growth.

Common errors

Error 1: assuming every square matrix can be diagonalized

It seems reasonable because the procedure to find eigenvalues is the same in every case. It fails because some matrices are defective: they have an eigenvalue whose algebraic multiplicity exceeds its geometric multiplicity, so they do not provide enough independent eigenvectors. Detect it by comparing the number of linearly independent eigenvectors obtained with the size of the matrix. Correct it by recognizing that such matrices require an extended basis of generalized eigenvectors instead of simple diagonalization. Mini-example: a matrix with a non-diagonal Jordan block produces fewer eigenvectors than its dimension.

Error 2: forgetting that an eigenvector is defined only up to scale

It seems reasonable to expect a unique answer for each eigenvector. It fails because any nonzero multiple of an eigenvector is still an eigenvector for the same eigenvalue. Detect it when two apparently different solutions turn out to be proportional. Correct it by normalizing by convention or leaving the free parameter explicit. Mini-example: (1, 2) and (2, 4) are the same eigenvector expressed at a different scale.

Error 3: confusing the order of the eigenvalues with their physical importance

It seems reasonable to simply sort eigenvalues from largest to smallest. It fails because in some problems the physically relevant eigenvalue is the smallest one, such as the fundamental oscillation mode, and in others it is the largest, such as the direction of maximum variance. Detect it by asking what each eigenvalue represents physically before sorting. Correct it by interpreting the physical meaning of each one in context, not only its magnitude. Mini-example: in normal modes, the smallest eigenvalue usually gives the fundamental frequency, not the mode dominant in energy.

Error 4: ignoring complex eigenvalues when they appear

It seems reasonable to discard complex solutions as unphysical. It fails because in systems with rotation, damping, or oscillation, complex eigenvalues are the norm and carry real physical information in both their real and imaginary parts. Detect it by checking whether the matrix is symmetric, which guarantees real eigenvalues, or not. Correct it by interpreting the real part as growth or decay and the imaginary part as oscillation frequency. Mini-example: a pair of complex conjugate eigenvalues in a dynamical system indicates a spiral, not a simple fixed point.

Error 5: substituting incorrectly into (A − λI)v = 0

It seems reasonable to subtract λ from just one entry of the matrix. It fails because λ must be subtracted only from the entries on the main diagonal, since the identity matrix multiplied by λ is being subtracted, not a loose scalar. Detect it by checking that the resulting matrix is still square and that only the diagonal changed. Correct it by explicitly writing A − λI before substituting the eigenvalue. Mini-example: subtracting λ from an off-diagonal entry invalidates the entire downstream calculation.

Quick detection rule

If the number of independent eigenvectors is smaller than the size of the matrix, if two "different" eigenvectors turn out to be proportional, or if a complex eigenvalue is dismissed without checking whether the problem allows it, the error is probably in the setup, not in the final arithmetic.

Physical applications

1. Coupled oscillations

It appears when describing systems of masses and springs connected to each other. The eigenvalues of the coupling matrix are calculated to obtain the squared natural frequencies. The result gives the normal modes of vibration. The dangerous error is trying to solve the equations of motion coordinate by coordinate instead of diagonalizing the system. As a final control, the learner must explain which mode is slower, which is faster, and how the masses move relative to each other in each one.

2. Rigid-body rotation

It appears in the inertia tensor, which relates angular momentum to angular velocity. Its eigenvectors are calculated to find the principal axes of inertia, directions where rotation produces no additional torque. The result greatly simplifies Euler's equations. The dangerous error is using arbitrary axes instead of the principal ones, which introduces unnecessary cross terms. As a final control, the learner must explain why rotating about a principal axis is more stable than rotating about an intermediate one.

3. Quantum mechanics

It appears because physical observables are represented by operators, and the possible values of a measurement are the eigenvalues of that operator. The eigenvectors are the eigenstates associated with each measurable value. The result connects linear algebra with experimental predictions. The dangerous error is forgetting that distinct eigenvalues of a Hermitian operator correspond to orthogonal states, which directly affects probability calculations. As a final control, the learner must explain what it physically means to measure a specific eigenvalue.

4. Stability of dynamical systems

It appears when linearizing a system of differential equations near an equilibrium point. The sign of the real part of the eigenvalues of the Jacobian matrix determines whether perturbations grow, decay, or oscillate. The result classifies the equilibrium as stable, unstable, or a saddle. The dangerous error is forgetting the imaginary part when it exists, which indicates oscillation in addition to growth or decay. As a final control, the learner must explain what happens to a small perturbation after a long time.

5. Data analysis and principal components

It appears when diagonalizing a covariance matrix to find the directions of maximum variability in a set of experimental data. The eigenvectors point along those directions and the eigenvalues indicate how much variance each one explains. The result allows the dimension of a problem to be reduced without losing relevant physical information. The dangerous error is dismissing a direction with a small eigenvalue as irrelevant without checking its physical meaning. As a final control, the learner must explain what fraction of the total variability each principal eigenvector captures.

Practice

Level 1: recognition

Identify, for a given matrix, what the steps to find its eigenvalues would be, without solving yet. State the size of the matrix, the expected degree of the characteristic polynomial, and whether you expect real or complex eigenvalues based on its symmetry.

Exercise: take a symmetric 2×2 matrix and a non-symmetric 2×2 matrix, and predict before calculating whether both will have real eigenvalues.

Level 2: direct application

Calculate the eigenvalues and eigenvectors of a simple 2×2 matrix with distinct real eigenvalues. Verify each pair by substituting into Av = λv.

Exercise: use an obviously block-diagonal matrix and check that its eigenvalues are directly the diagonal entries.

Level 3: choosing the tool

Decide whether to use the trace-and-determinant shortcut for a 2×2 matrix or cofactor expansion for a 3×3 matrix. Explain why the chosen strategy is more efficient in each case.

Exercise: set up the characteristic polynomial of a 3×3 matrix and decide whether it is worth searching for an integer root first before applying the general formula.

Level 4: physical interpretation

Interpret a set of eigenvalues and eigenvectors as the normal modes of a mass-spring system, or as the principal axes of a rigid body. Explain what the largest and smallest eigenvalues represent physically.

Exercise: given a pair of eigenvalues and eigenvectors of a two-mass coupled system, describe in words how each mass moves in each mode.

Level 5: intentional error

Analyze an eigenvalue calculation where λ was subtracted incorrectly or where an unverified eigenvector was accepted. Detect the faulty step, correct it, and explain what physical consequence the error would have had if it had gone undetected.

Exercise: invent a plausible error in setting up (A − λI)v = 0, write why it seems reasonable, correct it, and state which incorrect eigenvector it would have produced.