Eigenvalues & Eigenvectors

60 min12 pages

What is Eigenvalues & Eigenvectors?

Special directions preserved by transformations and their scaling factors. Key to principal component analysis, quantum mechanics, and system stability.

~60 min12 pages
eigenvalueseigenvectorstransformations

Eigenvalues and eigenvectors are fundamental concepts in linear algebra that describe how a linear transformation acts on space. An eigenvector is a nonzero vector that, when a linear transformation is applied, is only scaled by a factor called the eigenvalue, preserving its direction. This idea captures the intrinsic directions of a system that remain stable under the transformation. In many practical settings, such as computer graphics, physics, and data analysis, identifying these special directions helps simplify complex behavior. For example, rotating a vector by a matrix may leave certain directions unchanged in orientation, while simply stretching or shrinking their length by a factor. In two dimensions, consider a rotation-shear combination; some vectors align with invariant axes whose directions scale predictably. The eigen decomposition, when possible, expresses a matrix as PDP^{-1}, where D is diagonal with eigenvalues and P contains corresponding eigenvectors. This decomposition is powerful because diagonal matrices are easy to analyze and exponentiate, enabling solutions to differential equations and dynamics to be carried out efficiently. Understanding how to compute eigenvalues involves solving the characteristic equation det(A-λI)=0, which yields polynomial equations whose roots are the eigenvalues. Once eigenvalues are known, solving (A-λI)v=0 gives the eigenvectors associated with each eigenvalue. Although some matrices lack a full set of eigenvectors (defective), many practical matrices do provide complete eigensystems that unlock deep insights into stability, resonance, and modal behavior.

What is the eigenvector of a matrix A with eigenvalue λ?

It's any nonzero vector v such that Av = λv
It's any vector v such that Av = v/λ
It's a vector v orthogonal to Av
It's the vector that makes det(A-λI) nonzero

Sign up to unlock quizzes

Example: Simple 2x2 Eigenvalue Problem

Let A = [[2,0],[0,3]]. Its eigenvalues are λ1=2 and λ2=3 with corresponding eigenvectors v1=[1,0]^T and v2=[0,1]^T. Applying A to v1 yields Av1 = [2,0]^T = 2v1, and Av2 = [0,3]^T = 3v2. The matrix is already diagonal, so the eigenvectors align with the standard basis, and the eigenbasis diagonalizes A. This example shows how eigenvectors identify invariant directions under A.

Eigenvalues encode the scaling effect along eigenvectors when a linear transformation is applied. If a transformation represented by matrix A acts on a vector x, the result can often be decomposed into a sum of contributions along directions that behave independently. The eigenpairs (λ, v) satisfy Av = λv, meaning the direction v is preserved (up to scaling by λ). This concept is central to stability analysis in dynamical systems, where eigenvalues with magnitude greater than 1 imply divergence along a mode, while those with magnitude less than 1 indicate decay. In principal component analysis, data is projected onto eigenvectors of the covariance matrix; the eigenvalues indicate how much variance each principal component captures. In quantum mechanics, observables correspond to operators, whose eigenvectors form a basis of possible states with definite measurement outcomes, and eigenvalues correspond to the measured values. Moreover, the eigenspace associated with a given eigenvalue contains all vectors that scale by that eigenvalue under the transformation. When a matrix is symmetric, the eigenvectors corresponding to distinct eigenvalues are orthogonal, which simplifies both interpretation and computation. The process to compute eigenvalues usually starts with solving the characteristic polynomial det(A-λI)=0, a determinant that yields a polynomial equation in λ, whose roots are eigenvalues. After finding λs, one solves (A-λI)x=0 to obtain the eigenvectors. If the matrix is defective, fewer independent eigenvectors exist than the dimension of the space, complicating diagonalization but still amenable to Jordan form analysis.

The eigenvectors of a matrix A satisfy Av = _____, where _____ is the corresponding eigenvalue.

Type your answer...

Sign up to unlock quizzes

For a symmetric matrix A, eigenvectors corresponding to distinct eigenvalues are:

Always parallel
Orthogonal
Rotationally related
Identical to the standard basis

Sign up to unlock quizzes

Example: 2x2 Symmetric Matrix

text
A = [[4,1],[1,3]]
Compute eigenvalues by det(A-λI)=0:
|4-λ, 1; 1, 3-λ| = (4-λ)(3-λ) - 1 = λ^2 -7λ +11 = 0
Roots: λ = (7 ± sqrt(49-44))/2 = (7 ± sqrt(5))/2.
Corresponding eigenvectors from (A-λI)x=0 yield two independent vectors, e1 and e2, which are orthogonal due to symmetry.

Eigenvalues can be real or complex. For a real square matrix, eigenvalues may be complex conjugate pairs. The characteristic polynomial det(A-λI)=0 captures this spectrum. When A is diagonalizable, A = PDP^{-1} with D diagonal containing eigenvalues and P whose columns are eigenvectors. Diagonalization is immensely helpful: powers of A become A^k = PD^kP^{-1}, so repeated transformations are reduced to scaling along fixed directions. This is the backbone of linear dynamical systems, where state evolution x_{t+1} = Ax_t can be analyzed by decomposing into eigenmodes. In many applications, A is large; numerical methods such as QR algorithm or power iteration approximate dominant eigenpairs. The power method iteratively applies A to a vector to converge to the eigenvector associated with the largest magnitude eigenvalue, assuming certain conditions hold, like a spectral gap and a non-defective matrix. When dealing with non-diagonalizable matrices, one uses Jordan normal form, which still reveals generalized eigenvectors and chains. In data science, eigenvectors of covariance matrices point in directions of maximum variance; the top eigenvector is the first principal component, guiding dimensionality reduction. In mechanical systems, eigenvalues indicate natural frequencies; negative or complex eigenvalues can signal instability or oscillatory modes. The process of solving eigenproblems blends algebraic techniques with numerical methods, and understanding the geometry behind eigenpairs illuminates why these numbers govern so much of how linear systems behave.

Example: Power Method Outline

Given A with largest eigenvalue magnitude λ_max, start with x0 not orthogonal to the leading eigenvector. Iterate x_{k+1} = Ax_k / ||Ax_k||. As k grows, x_k converges to v_max, the eigenvector associated with λ_max, provided the eigenvalues are distinct in magnitude and the leading eigenvalue dominates others. This method is simple yet elegant for large sparse matrices.

What does the power method primarily compute for a matrix A with distinct eigenvalues by magnitude?

The smallest eigenvalue
The eigenvector corresponding to the eigenvalue with the largest magnitude
All eigenvectors simultaneously
The determinant of A

Sign up to unlock quizzes

In the eigenvalue equation Av = λv, if A is 2x2 and v = [x, y]^T, then solving (A-λI)v = 0 yields the eigenvectors corresponding to λ. The null space of (A-λI) is the eigenvector space, whose dimension equals the geometric multiplicity of λ.

Type your answer...

Sign up to unlock quizzes

This page introduces the geometric view: eigenvectors are invariant directions under linear transformation. If you visualize a matrix as a transformation in the plane, eigenvectors point along fixed lines that are not rotated, only stretched or compressed. The eigenvalue tells you how much scaling occurs along that line. For 2D matrices, you can often picture how different directions are stretched by factors λ1 and λ2. The spectral theorem guarantees that every real symmetric matrix is diagonalizable with an orthonormal basis of eigenvectors, making the action of the matrix simply scaling along orthogonal axes. When A is not symmetric or has complex eigenvalues, the geometry becomes richer: you may see rotation, shearing, or complex eigen-directions that correspond to oscillatory behavior in dynamics. To analyze a system, decompose an initial state into eigenvectors; each component evolves independently as λ^t times its eigenvector, allowing insight into stability and long-term behavior. The eigen-decomposition also underpins many numerical methods, including spectral clustering, where data is projected onto eigenvectors of Laplacians to reveal cluster structure. In programming, eigen-decomposition of a matrix is common in scientific computing libraries, with routines returning eigenvalues and eigenvectors that can be used for further analysis, such as solving differential equations or performing modal analysis in engineering. Mastery of the eigenvalue problem opens doors to interpreting linear transformations in both theoretical and applied contexts.

Example: Diagonalization Concept

If A = PDP^{-1} with D = diag(λ1, λ2,...), then A^k = PD^kP^{-1}. Each eigencomponent scales by λi^k along its eigenvector ei. This makes long-term behavior easy to analyze: components with |λi|<1 decay, |λi|>1 grow, and |λi|=1 persist or rotate if λi is a complex root of unity.

Which theorem guarantees an orthonormal basis of eigenvectors for every real symmetric matrix?

Spectral theorem (orthogonal diagonalization)
Pythagorean theorem
Rank-nullity theorem
Cayley-Hamilton theorem

Sign up to unlock quizzes

Eigenvalues can be complex; they come in conjugate pairs for real matrices when non-real. The complex eigenvalues manifest as oscillatory modes in dynamical systems: if an eigenvalue is a complex number a+bi with b≠0, the corresponding eigenvectors are complex and the real-valued system exhibits rotations and oscillations, often analyzed via real Jordan form or by pairing conjugate eigenvectors to form real invariant subspaces. The magnitude of a complex eigenvalue, |λ|, determines the damping or growth of the corresponding mode, while the angle (arg λ) determines rotation. In control theory, poles of a transfer function are eigenvalues of a companion or system matrix, and their locations in the complex plane diagnose stability and response speed. In quantum mechanics, Hermitian operators have real eigenvalues, and their eigenvectors form a complete basis for the Hilbert space, reflecting observable eigenstates with definite values. When working with data, eigenvectors of the covariance matrix point to directions of maximum variance; the magnitude of the eigenvalues indicates the amount of variance captured by each principal component. Numerical methods approximate eigenpairs for large matrices, with algorithms like QR decomposition iteratively refining estimates. The concept of eigenvalues extends beyond pure math into many disciplines, linking geometry, dynamics, statistics, and physics through the idea of invariant directions and scaling factors.

Example: Complex Eigenvalues in a 2x2 Real Matrix

text
A = [[0, -1],[1, 0]]
Eigenvalues are λ = i and λ = -i, with eigenvectors v = [1, i]^T and [1, -i]^T. Real systems exhibit rotation without growth since |λ|=1. If a small perturbation yields A = [[0, -1],[1, 0.1]], eigenvalues become complex with magnitude sqrt(0.1^2+1) > 1, indicating a rotating mode with slight growth.

What does a complex eigenvalue indicate about a dynamical system?

Pure stable growth along a real direction
Oscillation with growth/decay in amplitude
Pure decay without rotation
No physical interpretation

Sign up to unlock quizzes

The computation of eigenvalues and eigenvectors is central in data science, physics, and engineering. Numerically, one often uses algorithms that are robust to rounding errors and scale with the size of the matrix. The QR algorithm is a classic method that iteratively performs QR decompositions to converge to an upper triangular matrix whose diagonal entries approximate eigenvalues. For large sparse matrices, iterative methods like power iteration (dominant eigenpair) or Lanczos methods focus on a subset of eigenpairs efficiently. In PCA, we take the covariance matrix of the data; its eigenvectors define principal directions, and their eigenvalues quantify explained variance. When data are centered, eigenvectors produce a rotation of the coordinate system to align with directions of maximum variance. In quantum mechanics, the eigenvectors of a Hermitian operator form an orthonormal basis of eigenstates with real eigenvalues, making measurements deterministic in the eigenbasis. Stability analysis uses eigenvalues of the Jacobian matrix: if any eigenvalue lies outside the unit circle, the fixed point is unstable along the corresponding eigenvector, guiding design adjustments. In control theory, eigenvalues of the system matrix determine step response and damping, influencing how to shape a system for desired performance. In practice, you might compute eigenpairs with software like NumPy in Python (np.linalg.eig) or MATLAB's eig function, which provide both eigenvalues and eigenvectors, enabling a wide range of downstream tasks such as solving differential equations, performing modal analysis, or generating orthogonal bases for projection methods.

Example: PCA Intuition

Given a data matrix X (n samples, d features), compute the covariance C = X^T X / (n-1). The eigenvectors of C give principal directions; the largest eigenvalue indicates the direction of maximum variance. Project each data point onto this eigenvector to get the first principal component: y_i = v^T x_i.

Which numerical method is commonly used to approximate the largest eigenvalue of a large sparse matrix?

Power iteration
LU decomposition
Gaussian elimination
Cholesky factorization

Sign up to unlock quizzes

You've previewed 6 of 12 pages

Sign up free to unlock the rest of this lesson and start tracking your progress.

6 more pages waiting:

  • Page 7
  • Page 8
  • Page 9
  • Page 10
  • +2 more...

Related Topics