You are on page 1of 6

Eigenvectors and eigenvalues

1

Eigenvectors and eigenvalues
Jeremy Orloff
Here is a short note on eigenvalues and eigenvectors. We will look almost exclusively at 2x2 matrices. These have almost all the features of bigger square matrices and they are computationally easy. Contents Etymology. Definition. Why eigenvectors are special. Computational approach. Diagonal matrices and decoupling. Intutitive meaning and modes. Population matrices. Etymology: Eigenvector is German for ’own vector’. I can sort-of see how this makes sense, but it is not very descriptive for me. Definition: For a square matrix M an eigenvector is a non-zero vector v that satisfies the equation M v = λ v for some number λ. (1) The number λ is called the eigenvalue corresponding to v. We will call equation (1) the eigenvector equation. Comments: 1. Using λ for the eigenvalue is a fairly common practice when looking at generic matrices. If the eigenvalue has a physical interpretation I’ll often use a corresponding letter. For example, in population matrices the eigenvalues are growthrates, so I’ll use the letter r. 2. Eigenvectors are not unique. That is, if v is an eigenvector with eigenvalue λ then so is 2 v, 3.14 v, indeed so is any scalar multiple of v.

7) 2 u2 O Au1 x u1 −1  v2 = Av2 / 5 Action of the matrix A on vectors Example 2: Any rotation in three dimensions is around some axis. Take u1 = 1 0 ⇒ Au1 = 6 1 . it is unmoved by A. v2 = 1 −1 then Av2 = v2 . 1 2 We will explore how A transforms vectors and what makes an eigenvector special. Now we apply our ’fact’ to the last equation above to conclude that λ is an eigenvalue if and only if det(A − λ I) = 0 (2) . I. eigenvectors lie on lines that are unmoved by A. Likewise. as always. let I be the identity matrix. Computational approach: In order to find the eigenvectors and eigenvalues we use the following basic fact about matrices. We manipulate the eigenvector equation (equation 1) as follows. A v = λ v ⇔ A v = λ Iv ⇔ A v − λ Iv = 0 ⇔ (A − λ I)v = 0.Eigenvectors and eigenvalues 2 Why eigenvectors are special: 6 5 Example 1: Let A = . Take u2 = 0 1 ⇒ Au2 = 5 2 . The vector along this axis is fixed by the rotation. The eigenvector v2 is really special. Now take v1 = 5 1 ⇒ Av1 = 35 7 = 7v1 . We will see that A scales and rotates most vectors. y 2 Au2 5 v2 1 to Av1 = (35. The eigenvector is special since A just scales it by 7. it is an eigenvector with eigenvalue 1.e. We see that A scales and turns most vectors. Fact: The equation M v = 0 is satisfied by a non-zero vector v if and only if det(M ) = 0. Now. but only scales eigenvectors. That is.

Example 4: Consider the diagonal matrix B = Convince yourself that B the y-coordinate by 3. any scalar multiple of these eigenvectors is also an eigenvector with the same eigenvalue.Eigenvectors and eigenvalues (Recall that an eigenvector is by definition non-zero. Take v = 1 −1 5 1 . Comparing this with the eigenvector equation (equation (1)) we see that .) It allows us to find the eigenvalues and eigenvectors separately in a two step process. That is B scales the x-coordinate by 2 and 1 0 and 0 1 are eigenvectors with eigenvalues 2 and 3 respectively. = 0 0 0 0 ⇒ −a1 + 5a2 = 0. Step 2. . Take v = ⇒ 5a1 + 5a2 = 0.) 3 We call equation (2) the characteristic equation. Notation: det(A) = |A|. 6 5 1 2 . (Eigenvalues are sometimes called characteristic values. ⇒ − =0 ⇒ 1 2−λ 1 2 0 λ ⇒ (6 − λ)(2 − λ) − 5 = λ2 − 8λ + 7 = 0 ⇒ λ = 7. a2 λ = 7: λ = 1: −1 5 1 −5 5 5 1 1 a1 a2 a1 a2 = (A − λI)v = 0. find λ (eigenvalues): |A − λI| = 0 (characteristic equation) 6−λ 5 6 5 λ 0 = 0. To repeat the comment above. Diagonal matrices and decoupling: In this section we will see how easy it is to work with diagonal matrices and how finding eigenvalues and eigenvectors of a matrix is like turning it into a diagonal matrix. 1. find v (eigenvectors): a1 Let v = . That is. for a diagonal matrix the diagonal entries are the eigenvalues and the eigenvectors all point along the coordinate axes. Example 3: Find the eigenvalues and eigenvectors of the matrix A = Step 1. We also have B 1 0 =2 1 0 and B 0 1 =3 0 1 x y = 2x 3y 2 0 0 3 . Comment: In Matlab the function ’eig’ returns the eigenvectors and eigenvalues of a matrix.

x(t + 1) = 6x(t) + 5y(t) y(t + 1) = x(t) + 2y(t). Suppose x and y are physical values that are updated at intervals according to the relations x(t + 1) = 2x(t) y(t + 1) = 3y(t). Thinking of x and y as populations let’s see what happens to the system from various starting points. There is of course one problem. = 1 −1 .Eigenvectors and eigenvalues 4 Before going on let’s see why we call this a decoupled system.5 0. =A x(0) y(0) = 1 −1 . That is. is the matrix in examples (1) and (3) above. We consider x and y to be decoupled because their values have no affect on each other. . We see that our initial population distribution is a stable one that stays constant over time.5 .5 y(0) Then x(1) y(1) Generally x(t) y(t) = At x(0) y(0) = 7t 2. In this system the variables x and y are coupled. (ii) Start from another eigenvector: Let x(0) y(0) Then x(1) y(1) Generally x(t) y(t) = At x(0) y(0) = 1 −1 . the initial vector has a negative value and is thus not a realistic population vector.5 .5): Let 2. (i) Start from an eigenvector (we scale the eigenvector in example (3) by . the eigenvector represents a stable distribution and the eigenvalue its growth rate. =A x(0) y(0) =7 2.5 x(0) . We see that our initial population distribution is a stable one that grows exponentially over time.5 0. = 0. Example 5: Now consider the system In matrix form this is x(t + 1) y(t + 1) where A = 6 5 1 2 =A x(t) y(t) .

we can decompose our initial vector into a sum of eigenvectors. This is a standard phenomenon. That is. We have x(t + 1) y(t + 1) =A x(t) y(t) = A u(t) 5 1 + v(t) 1 −1 . So. In order to make it explicit we introduce two new variables u and v with the relation x y =u 5 1 +v 1 −1 .Eigenvectors and eigenvalues (iii) Start from a non-eigenvector: Let x(0) y(0) Then x(1) y(1) =A x(0) y(0) = 69 15 . u and v are the coefficients in the decomposition of x into eigenvectors. Now it’s easy to apply A any number of times x(t) y(t) Notice that over time the = At x(0) y(0) = 7t 2 5 1 − 1 −1 . Unwinding this we respect to u and v we have the two equations x(t + 1) y(t + 1) and x(t + 1) y(t + 1) = 7 u(t) 5 1 + v(t) 1 −1 . The above hints at diagonalization and decoupling. Here’s where eigenvalues do their thing. 5 Working like this it is difficult to see how the population will evolve. = u(t + 1) 5 1 + v(t + 1) 1 −1 . 1 piece of the population vector becomes insignificant when −1 compared with the first piece. Note 9 3 =2 5 1 − 1 −1 . For the y moment we will set aside any physical interpretation of u and v and see how they act as a system. That is we have the decoupled system u(t + 1) = 7u(t) v(t + 1) = v(t) . the others are insignificant in the long run. the largest eigenvalue is the dominant one. = 9 3 .

Some eigenvectors may have negative entries and some eigenvalues may be negative or complex. The axes are eigenvectors of the coefficient matrix for the system of differential equations modeling this system. However. • • a1 • Population matrices: Here we still think of each of the eigenvectors as pure modes. 6 Example 6: A book has three axes of symmetry. Spinning in any other way is a combination of these pure modes. there is one realistic mode and it is dominant in the sense that the other modes become insignificant in the long run. any population vector is a combination of these pure modes. Intutitive meaning and modes: Physically we think of each of the eigenvectors as representing pure modes. .Eigenvectors and eigenvalues Note that the eigenvalues of A are precisely the coefficients of the decoupled system.) These are the pure modes. (Although the one labeled a1 is so unstable that it is difficult to actually spin the book around it. as in example (5). It will spin nicely about these axes. Again. As in example (5) not all the pure modes are physically realistic.