Linear AlgebraEigenvalues and Eigenvectors2 of 7
Eigenvalues and Eigenvectors · 2

Finding Eigenvalues: The Characteristic Equation

To find the eigenvalues and eigenvectors of a matrix AA, we need a way to solve Av=λvA\eigvec{v} = \eigval{\lambda} \eigvec{v} when the matrix AA is the only thing we know.

1

Building the Characteristic Equation

For a matrix like A=[3124]A = \begin{bmatrix} 3 & 1 \\ 2 & 4 \end{bmatrix}, most vectors get knocked off their original line when multiplied by AA, while an eigenvector v\eigvec{v} stays on its line and simply stretches or squishes by its eigenvalue λ\eigval{\lambda}:

Av\displaystyle A\eigvec{v} =λv\displaystyle = \eigval{\lambda} \eigvec{v}
Knocked off its line
uu
AuA u
Stays on its line
v\htmlClass{lc-blue}{v}
×λ\htmlClass{lc-violet}{\times\lambda}
Av=λvA\htmlClass{lc-blue}{v} = \htmlClass{lc-violet}{\lambda}\htmlClass{lc-blue}{v}

Right now, both the eigenvalue λ\eigval{\lambda} and the eigenvector v\eigvec{v} are unknown. To find λ\eigval{\lambda} on its own, we need a way to pull all the unknowns to one side and eliminate v\eigvec{v}.

Building the Characteristic Equation
uu
v\htmlClass{lc-blue}{v}
A=[3124]Av=λvλ=  ?v=  ?\displaystyle \begin{aligned} A &= \begin{bmatrix} 3 & 1 \\ 2 & 4 \end{bmatrix} \\[6pt] A\htmlClass{lc-blue}{v} &= \eigval{\lambda}\htmlClass{lc-blue}{v} \\[6pt] \eigval{\lambda} &= \;?\qquad \htmlClass{lc-blue}{v} = \;? \end{aligned}

To factor v\eigvec{v} out of Av−λv=[00]A\eigvec{v} - \eigval{\lambda} \eigvec{v} = \begin{bmatrix} 0 \\ 0 \end{bmatrix}, we first rewrite the number λ\eigval{\lambda} as the matrix λI=[λ00λ]\eigval{\lambda} I = \begin{bmatrix} \eigval{\lambda} & 0 \\ 0 & \eigval{\lambda} \end{bmatrix}, which scales any vector by λ\eigval{\lambda} just like the number λ\eigval{\lambda} does:

Av=λv⟶Av−λIv=[00]⟶(A−λI)v=[00]\displaystyle \begin{aligned} A\eigvec{v} &= \eigval{\lambda} \eigvec{v} \\ \quad\longrightarrow\quad A\eigvec{v} - \eigval{\lambda} I \eigvec{v} &= \begin{bmatrix} 0 \\ 0 \end{bmatrix} \\ \quad\longrightarrow\quad (A - \eigval{\lambda} I)\eigvec{v} &= \begin{bmatrix} 0 \\ 0 \end{bmatrix} \end{aligned}

Writing out the entries, A−λIA - \eigval{\lambda} I simply subtracts λ\eigval{\lambda} from the main diagonal of AA:

[3−λ124−λ][xy]=[00]\displaystyle \begin{bmatrix} 3 - \eigval{\lambda} & 1 \\ 2 & 4 - \eigval{\lambda} \end{bmatrix}\eigvec{\begin{bmatrix} x \\ y \end{bmatrix}} = \begin{bmatrix} 0 \\ 0 \end{bmatrix}

Because an eigenvector v\eigvec{v} cannot be [00]\begin{bmatrix} 0 \\ 0 \end{bmatrix}, the matrix A−λIA - \eigval{\lambda} I must send a non-zero vector to [00]\begin{bmatrix} 0 \\ 0 \end{bmatrix}. Remember, a matrix sends a non-zero vector to [00]\begin{bmatrix} 0 \\ 0 \end{bmatrix} only when it squishes space flat, which means its determinant is zero:

det⁡(A−λI)\displaystyle \det(A - \eigval{\lambda} I) =0\displaystyle = 0

As λ\eigval{\lambda} changes the main diagonal entries of A−λIA - \eigval{\lambda} I, the transformed grid stays two-dimensional everywhere except at two exact values of λ\eigval{\lambda}, where the grid squishes flat onto a single line (det⁡(A−λI)=0\det(A - \eigval{\lambda} I) = 0) and a whole line of non-zero vectors collapses to [00]\begin{bmatrix} 0 \\ 0 \end{bmatrix}.

Shifting the diagonal sends non-zero vectors to the origin only at the two shift amounts that squish the grid flat onto a line.
2

Solving the Characteristic Equation for a 2×22\times 2 Matrix

Now let's find the eigenvalues of A=[3124]A = \begin{bmatrix} 3 & 1 \\ 2 & 4 \end{bmatrix} by solving its characteristic equation det⁡(A−λI)=0\det(A - \eigval{\lambda} I) = 0, and then plug an eigenvalue back into (A−λI)v=[00](A - \eigval{\lambda} I)\eigvec{v} = \begin{bmatrix} 0 \\ 0 \end{bmatrix} to find its eigenvector.

Solving the Characteristic Equation for a 2×2 Matrix
A=[3124]det⁡(A−λI)=0λ=  ?\displaystyle \begin{aligned} A &= \begin{bmatrix} 3 & 1 \\ 2 & 4 \end{bmatrix} \\[4pt] \det(A - \eigval{\lambda} I) &= 0 \\[4pt] \eigval{\lambda} &= \;? \end{aligned}
Text solution
Subtract λ\eigval{\lambda} from the main diagonal and set the determinant to zeroWrite out the 2×22\times 2 determinant of A−λIA - \eigval{\lambda} I by multiplying the main diagonal entries and subtracting the product of the other two entries:
det⁡[3−λ124−λ]=(3−λ)(4−λ)−(1)(2)=0⏟characteristic equation\displaystyle \det\begin{bmatrix} 3 - \eigval{\lambda} & 1 \\ 2 & 4 - \eigval{\lambda} \end{bmatrix} = \underbrace{(3 - \eigval{\lambda})(4 - \eigval{\lambda}) - (1)(2) = 0}_{\text{characteristic equation}}
Expand and factor the characteristic polynomialMultiply out the terms, combine like terms into a quadratic in λ\eigval{\lambda}, and factor to find the eigenvalues:
12−3λ−4λ+λ2−2=0λ2−7λ+10⏟characteristic polynomial=0(λ−5)(λ−2)=0⟶λ=5andλ=2\displaystyle \begin{aligned} 12 - 3\eigval{\lambda} - 4\eigval{\lambda} + \eigval{\lambda}^2 - 2 &= 0 \\ \underbrace{\eigval{\lambda}^2 - 7\eigval{\lambda} + 10}_{\text{characteristic polynomial}} &= 0 \\ (\eigval{\lambda} - 5)(\eigval{\lambda} - 2) &= 0 \\ \longrightarrow\quad \eigval{\lambda} &= \eigval{5} \\ \text{and}\quad \eigval{\lambda} &= \eigval{2} \end{aligned}
Plug λ=5\eigval{\lambda} = \eigval{5} into (A−λI)v=[00](A - \eigval{\lambda} I)\eigvec{v} = \begin{bmatrix} 0 \\ 0 \end{bmatrix} to find its eigenvectorSubstitute λ=5\eigval{\lambda} = \eigval{5} into A−λIA - \eigval{\lambda} I and solve the top row for a non-zero vector [xy]\eigvec{\begin{bmatrix} x \\ y \end{bmatrix}}:
[3−5124−5][xy]=[−212−1][xy]=[00]⟶−2x+1y=0⟶y=2x\displaystyle \begin{aligned} \begin{bmatrix} 3 - \eigval{5} & 1 \\ 2 & 4 - \eigval{5} \end{bmatrix}\eigvec{\begin{bmatrix} x \\ y \end{bmatrix}} & \\ = \begin{bmatrix} -2 & 1 \\ 2 & -1 \end{bmatrix}\eigvec{\begin{bmatrix} x \\ y \end{bmatrix}} &= \begin{bmatrix} 0 \\ 0 \end{bmatrix} \\ \longrightarrow\quad -2\eigvec{x} + 1\eigvec{y} &= 0 \\ \longrightarrow\quad \eigvec{y} &= 2\eigvec{x} \end{aligned}
Pick a non-zero vector on the line and check Av=5vA\eigvec{v} = \eigval{5}\eigvec{v}Choosing x=1\eigvec{x} = 1 gives y=2\eigvec{y} = 2, so v=[12]\eigvec{v} = \eigvec{\begin{bmatrix} 1 \\ 2 \end{bmatrix}} is an eigenvector for λ=5\eigval{\lambda} = \eigval{5}, and multiplying by AA stretches it by 5\eigval{5} along the line y=2x\eigvec{y} = 2\eigvec{x}:
A[12]=[(3)(1)+(1)(2)(2)(1)+(4)(2)]=[510]=5[12]\displaystyle \begin{aligned} A\eigvec{\begin{bmatrix} 1 \\ 2 \end{bmatrix}} &= \begin{bmatrix} (3)(\eigvec{1}) + (1)(\eigvec{2}) \\ (2)(\eigvec{1}) + (4)(\eigvec{2}) \end{bmatrix} \\ &= \begin{bmatrix} 5 \\ 10 \end{bmatrix} = \eigval{5}\eigvec{\begin{bmatrix} 1 \\ 2 \end{bmatrix}} \end{aligned}
11
55
22
1010
v=[12]\htmlClass{lc-blue}{v} = \htmlClass{lc-blue}{\begin{bmatrix} 1 \\ 2 \end{bmatrix}}
×5\htmlClass{lc-violet}{\times 5}
y=2x\htmlClass{lc-blue}{y} = 2\htmlClass{lc-blue}{x}
Av=5vA\htmlClass{lc-blue}{v} = \htmlClass{lc-violet}{5}\htmlClass{lc-blue}{v}

Not every matrix has real eigenvalues: if a matrix rotates every vector off its original line (like a 90∘90^\circ rotation), its characteristic polynomial has no real roots.

3

Trace, Determinant, and Zero Eigenvalues

The sum of the entries on the main diagonal of a matrix is called the trace of the matrix, written Trace⁡(A)\diagsum{\operatorname{Trace}(A)}. Together, Trace⁡(A)\diagsum{\operatorname{Trace}(A)} and det⁡(A)\detval{\det(A)} connect the eigenvalues directly back to the entries of AA and tell us what happens when λ=0\eigval{\lambda} = \eigval{0} is an eigenvalue.

Trace, Determinant, and Zero Eigenvalues
A=[3124]λ=5, 2\displaystyle A = \begin{bmatrix} 3 & 1 \\ 2 & 4 \end{bmatrix} \qquad \eigval{\lambda} = \eigval{5},\ \eigval{2}
3+4=  ?5+2=  ?det⁡(A)=  ?5×2=  ?\displaystyle \begin{alignedat}{2} \diagsum{3 + 4} &= \;? &\qquad \eigval{5} + \eigval{2} &= \;? \\[4pt] \detval{\det(A)} &= \;? &\qquad \eigval{5} \times \eigval{2} &= \;? \end{alignedat}

Expanding det⁡(A−λI)=0\det(A - \eigval{\lambda} I) = 0 for a general 2×22\times 2 matrix A=[abcd]A = \begin{bmatrix} a & b \\ c & d \end{bmatrix} and matching it with the factored form (λ−λ1)(λ−λ2)=0(\eigval{\lambda} - \eigval{\lambda_1})(\eigval{\lambda} - \eigval{\lambda_2}) = 0 shows why the sum and product of the eigenvalues show up right in the coefficients:

(a−λ)(d−λ)−bc=λ2−(a+d)⏟Trace⁡(A)λ+(ad−bc)⏟det⁡(A)=0(λ−λ1)(λ−λ2)=λ2−(λ1+λ2)λ+λ1λ2=0\displaystyle \begin{aligned} &(a - \eigval{\lambda})(d - \eigval{\lambda}) - bc \\ &= \eigval{\lambda}^2 - \diagsum{\underbrace{(a + d)}_{\operatorname{Trace}(A)}}\eigval{\lambda} + \detval{\underbrace{(ad - bc)}_{\det(A)}} = 0 \\[6pt] &(\eigval{\lambda} - \eigval{\lambda_1})(\eigval{\lambda} - \eigval{\lambda_2}) \\ &= \eigval{\lambda}^2 - (\eigval{\lambda_1} + \eigval{\lambda_2})\eigval{\lambda} + \eigval{\lambda_1}\eigval{\lambda_2} = 0 \end{aligned}

Matching the coefficients gives two rules that work as a quick check on any problem:

Sum of eigenvalues=Trace⁡(A),Product of eigenvalues=det⁡(A)\displaystyle \begin{aligned} \text{Sum of eigenvalues} &= \diagsum{\operatorname{Trace}(A)}, \\ \text{Product of eigenvalues} &= \detval{\det(A)} \end{aligned}

While an eigenvector v\eigvec{v} can never be [00]\begin{bmatrix} 0 \\ 0 \end{bmatrix}, an eigenvalue can be λ=0\eigval{\lambda} = \eigval{0}. When λ=0\eigval{\lambda} = \eigval{0}, the equation Av=0v=[00]A\eigvec{v} = \eigval{0}\eigvec{v} = \begin{bmatrix} 0 \\ 0 \end{bmatrix} says that the original matrix AA sends a non-zero vector straight to [00]\begin{bmatrix} 0 \\ 0 \end{bmatrix} on its own, which means AA squishes space flat and the product of its eigenvalues is zero:

λ=0 is an eigenvalue of A⟺det⁡(A)=0⟺A is not invertible\displaystyle \begin{aligned} &\eigval{\lambda} = \eigval{0} \text{ is an eigenvalue of } A \\ \Longleftrightarrow\quad &\detval{\det(A)} = 0 \\ \Longleftrightarrow\quad &A \text{ is not invertible} \end{aligned}

Along the two eigenvector lines, stretching by λ1\eigval{\lambda_1} in one direction and λ2\eigval{\lambda_2} in the other keeps the main diagonal sum Trace⁡(A)\diagsum{\operatorname{Trace}(A)} equal to λ1+λ2\eigval{\lambda_1} + \eigval{\lambda_2} and scales the area by det⁡(A)=λ1λ2\detval{\det(A)} = \eigval{\lambda_1}\eigval{\lambda_2}. When either stretch factor is λ=0\eigval{\lambda} = \eigval{0}, that direction collapses to [00]\begin{bmatrix} 0 \\ 0 \end{bmatrix} and space flattens to zero area (det⁡(A)=0\detval{\det(A)} = 0).

The two eigenvalues (how much the matrix stretches along its two eigenvector lines) always add up to the matrix's diagonal sum (the trace) and multiply to its area scale factor (the determinant).
4

Eigenvalues of Triangular Matrices

For a triangular matrix, where every entry below the main diagonal is zero, or every entry above the main diagonal is zero, we can read off the eigenvalues without expanding a polynomial.

Eigenvalues of Triangular Matrices
A=[640−2]λ=  ?\displaystyle \begin{aligned} A &= \begin{bmatrix} 6 & 4 \\ 0 & -2 \end{bmatrix} \\[4pt] \eigval{\lambda} &= \;? \end{aligned}

When we subtract λ\eigval{\lambda} from the main diagonal of an upper triangular matrix like [640−2]\begin{bmatrix} 6 & 4 \\ 0 & -2 \end{bmatrix}, the zero below the diagonal makes the second part of the determinant zero, leaving the characteristic equation already factored:

det⁡[6−λ40−2−λ]=(6−λ)(−2−λ)−(4)(0)=(6−λ)(−2−λ)=0⟶λ=6 and λ=−2\displaystyle \begin{aligned} &\det\begin{bmatrix} 6 - \eigval{\lambda} & 4 \\ 0 & -2 - \eigval{\lambda} \end{bmatrix} \\ &= (6 - \eigval{\lambda})(-2 - \eigval{\lambda}) - (4)(0) \\ &= (6 - \eigval{\lambda})(-2 - \eigval{\lambda}) = 0 \\ &\longrightarrow\quad \eigval{\lambda} = \eigval{6} \text{ and } \eigval{\lambda} = \eigval{-2} \end{aligned}
5

Finding Eigenvalues of a 3×33\times 3 Matrix

When a 3×33\times 3 matrix is not triangular, such as A=[010−1313−13]A = \begin{bmatrix} 0 & 1 & 0 \\ -1 & 3 & 1 \\ 3 & -1 & 3 \end{bmatrix}, we find its eigenvalues by setting det⁡(A−λI)=0\det(A - \eigval{\lambda} I) = 0, where A−λIA - \eigval{\lambda} I squishes 3D volume down to zero.

To take the 3×33\times 3 determinant, we expand along a row or column, ideally one with a zero so one of the three 2×22\times 2 determinants drops out.

Finding Eigenvalues of a 3×3 Matrix
A=[010−1313−13]λ=  ?\displaystyle \begin{aligned} A &= \begin{bmatrix} 0 & 1 & 0 \\ -1 & 3 & 1 \\ 3 & -1 & 3 \end{bmatrix} \\[6pt] \eigval{\lambda} &= \;? \end{aligned}
Text solution
Subtract λ\eigval{\lambda} from the main diagonal and expand along the top rowApply the top-row sign pattern +,−,++, -, + so the 00 in the third column drops out:
det⁡[−λ10−13−λ13−13−λ]=(−λ)det⁡[3−λ1−13−λ]−1⋅det⁡[−1133−λ]+0=(−λ)((3−λ)(3−λ)−(1)(−1))−1((−1)(3−λ)−(1)(3))=0\displaystyle \begin{aligned} &\det\begin{bmatrix} -\eigval{\lambda} & 1 & 0 \\ -1 & 3 - \eigval{\lambda} & 1 \\ 3 & -1 & 3 - \eigval{\lambda} \end{bmatrix} \\ &= (-\eigval{\lambda})\det\begin{bmatrix} 3 - \eigval{\lambda} & 1 \\ -1 & 3 - \eigval{\lambda} \end{bmatrix} \\ &\quad - 1\cdot\det\begin{bmatrix} -1 & 1 \\ 3 & 3 - \eigval{\lambda} \end{bmatrix} + 0 \\ &= (-\eigval{\lambda})\big((3 - \eigval{\lambda})(3 - \eigval{\lambda}) \\ &\qquad\quad - (1)(-1)\big) \\ &\quad - 1\big((-1)(3 - \eigval{\lambda}) - (1)(3)\big) = 0 \end{aligned}
Expand into a cubic polynomial in λ\eigval{\lambda}Work out the two 2×22\times 2 determinants, multiply through, and multiply both sides by −1-1 so the λ3\eigval{\lambda}^3 term is positive:
(−λ)(λ2−6λ+10)−1(λ−6)=0−λ3+6λ2−10λ−λ+6=0−λ3+6λ2−11λ+6=0λ3−6λ2+11λ−6=0\displaystyle \begin{aligned} (-\eigval{\lambda})(\eigval{\lambda}^2 - 6\eigval{\lambda} + 10)& \\ {}- 1(\eigval{\lambda} - 6) &= 0 \\ -\eigval{\lambda}^3 + 6\eigval{\lambda}^2 - 10\eigval{\lambda} - \eigval{\lambda} + 6 &= 0 \\ -\eigval{\lambda}^3 + 6\eigval{\lambda}^2 - 11\eigval{\lambda} + 6 &= 0 \\ \eigval{\lambda}^3 - 6\eigval{\lambda}^2 + 11\eigval{\lambda} - 6 &= 0 \end{aligned}
Find an integer root and factor the cubicTest λ=1\eigval{\lambda} = \eigval{1} by adding the coefficients: (1)3−6(1)2+11(1)−6(\eigval{1})^3 - 6(\eigval{1})^2 + 11(\eigval{1}) - 6 =1−6+11−6=0{}= 1 - 6 + 11 - 6 = 0, so (λ−1)(\eigval{\lambda} - 1) is a factor. Factor out (λ−1)(\eigval{\lambda} - 1) and then factor the remaining quadratic:
(λ−1)(λ2−5λ+6)=0(λ−1)(λ−2)(λ−3)=0⟶λ=1,  λ=2,  λ=3\displaystyle \begin{aligned} (\eigval{\lambda} - 1)(\eigval{\lambda}^2 - 5\eigval{\lambda} + 6) &= 0 \\ (\eigval{\lambda} - 1)(\eigval{\lambda} - 2)(\eigval{\lambda} - 3) &= 0 \\ \longrightarrow\quad \eigval{\lambda} &= \eigval{1},\; \\ \eigval{\lambda} &= \eigval{2},\; \\ \eigval{\lambda} &= \eigval{3} \end{aligned}
Check the eigenvalues with the traceConfirm that the three eigenvalues add up to the sum of the main diagonal of AA:
1+2+3\displaystyle \eigval{1} + \eigval{2} + \eigval{3} =6\displaystyle = 6 =0+3+3\displaystyle = \diagsum{0} + \diagsum{3} + \diagsum{3} =Trace⁡(A)\displaystyle = \diagsum{\operatorname{Trace}(A)}
6

Similar Matrices Have the Same Eigenvalues

Two matrices AA and BB are called similar when B=P−1APB = P^{-1}AP, which means AA and BB describe the exact same transformation in two different coordinate systems converted by PP and P−1P^{-1}.

Similar Matrices Have the Same Eigenvalues
B=P−1APeigenvalues of B=  ?\displaystyle \begin{aligned} B &= P^{-1}AP \\[6pt] \text{eigenvalues of } B &= \;? \end{aligned}
v\htmlClass{lc-blue}{v}
×5\htmlClass{lc-violet}{\times 5}
AvA\htmlClass{lc-blue}{v}

To compare the characteristic polynomial of B=P−1APB = P^{-1}AP with that of AA, we use two facts: writing the identity matrix in a new coordinate system leaves it unchanged (P−1IP=P−1P=IP^{-1}IP = P^{-1}P = I, so λI=P−1(λI)P\eigval{\lambda} I = P^{-1}(\eigval{\lambda} I)P), and the determinant of a product of matrices is the product of their determinants:

det⁡(B−λI)=det⁡(P−1AP−λI)=det⁡(P−1AP−P−1(λI)P)=det⁡(P−1(A−λI)P)=det⁡(P−1)⋅det⁡(A−λI)⋅det⁡(P)=det⁡(A−λI)since det⁡(P−1)⋅det⁡(P)=1\displaystyle \begin{aligned} &\det(B - \eigval{\lambda} I) \\ &= \det(P^{-1}AP - \eigval{\lambda} I) \\ &= \det\big(P^{-1}AP - P^{-1}(\eigval{\lambda} I)P\big) \\ &= \det\big(P^{-1}(A - \eigval{\lambda} I)P\big) \\ &= \det(P^{-1})\cdot\det(A - \eigval{\lambda} I)\cdot\det(P) \\ &= \det(A - \eigval{\lambda} I) \\ &\qquad \text{since } \det(P^{-1})\cdot\det(P) = 1 \end{aligned}

As the basis PP changes, every entry of B=P−1APB = P^{-1}AP updates to describe the transformation in the new grid, while the stretched eigenvector lines in the plane, the characteristic polynomial det⁡(B−λI)\det(B - \eigval{\lambda} I), Trace⁡(B)\diagsum{\operatorname{Trace}(B)}, det⁡(B)\detval{\det(B)}, and the eigenvalues stay completely unchanged.

Changing the coordinate grid changes every entry of the matrix, while its characteristic polynomial, trace, determinant, and eigenvalues stay at the exact same values.

Key concepts

Before expanding a determinant to find eigenvalues, check whether you can read them right off the matrix. If the matrix is triangular (all zeros below or above the main diagonal, like [640−2]\begin{bmatrix} 6 & 4 \\ 0 & -2 \end{bmatrix}), the second part of the determinant is zero and the eigenvalues sit right on the main diagonal (λ=6\eigval{\lambda} = \eigval{6} and λ=−2\eigval{\lambda} = \eigval{-2}). If the matrix is similar to a matrix AA you already know (B=P−1APB = P^{-1}AP), they represent the same transformation in different coordinate systems, so BB has the exact same characteristic polynomial, eigenvalues, trace, and determinant as AA.

When you do need to solve for the eigenvalues from scratch, subtract λ\eigval{\lambda} from the main diagonal and set the determinant to zero so the matrix A−λIA - \eigval{\lambda} I squishes space flat:

det⁡(A−λI)\displaystyle \det(A - \eigval{\lambda} I) =0\displaystyle = 0

For a 2×22\times 2 matrix like A=[3124]A = \begin{bmatrix} 3 & 1 \\ 2 & 4 \end{bmatrix}, setting (3−λ)(4−λ)−(1)(2)=0(3 - \eigval{\lambda})(4 - \eigval{\lambda}) - (1)(2) = 0 gives the quadratic characteristic polynomial λ2−7λ+10=(λ−5)(λ−2)=0\eigval{\lambda}^2 - 7\eigval{\lambda} + 10 = (\eigval{\lambda} - 5)(\eigval{\lambda} - 2) = 0, so λ=5\eigval{\lambda} = \eigval{5} and λ=2\eigval{\lambda} = \eigval{2} (and if that quadratic has no real roots, the matrix rotates every vector off its line and has no real eigenvalues). For a 3×33\times 3 matrix like [010−1313−13]\begin{bmatrix} 0 & 1 & 0 \\ -1 & 3 & 1 \\ 3 & -1 & 3 \end{bmatrix}, expand det⁡(A−λI)=0\det(A - \eigval{\lambda} I) = 0 along the row with the most zeros to get a cubic (λ3−6λ2+11λ−6=0\eigval{\lambda}^3 - 6\eigval{\lambda}^2 + 11\eigval{\lambda} - 6 = 0), test λ=1\eigval{\lambda} = \eigval{1} and small integer factors of the constant term to pull out the first factor (λ−1)(\eigval{\lambda} - 1), and factor the remaining quadratic. Once you have an eigenvalue λ\eigval{\lambda}, plug it back into (A−λI)v=[00](A - \eigval{\lambda} I)\eigvec{v} = \begin{bmatrix} 0 \\ 0 \end{bmatrix} (such as [−212−1][xy]=[00]\begin{bmatrix} -2 & 1 \\ 2 & -1 \end{bmatrix}\eigvec{\begin{bmatrix} x \\ y \end{bmatrix}} = \begin{bmatrix} 0 \\ 0 \end{bmatrix} for λ=5\eigval{\lambda} = \eigval{5}) and pick any non-zero vector on the resulting line (v=[12]\eigvec{v} = \eigvec{\begin{bmatrix} 1 \\ 2 \end{bmatrix}}).

Finally, you can always check your eigenvalues, or find a missing one, straight from the original matrix AA:

Sum of eigenvalues=Trace⁡(A),Product of eigenvalues=det⁡(A)\displaystyle \begin{aligned} \text{Sum of eigenvalues} &= \diagsum{\operatorname{Trace}(A)}, \\ \text{Product of eigenvalues} &= \detval{\det(A)} \end{aligned}

For A=[3124]A = \begin{bmatrix} 3 & 1 \\ 2 & 4 \end{bmatrix}, the eigenvalues 5\eigval{5} and 2\eigval{2} add to Trace⁡(A)=3+4=7\diagsum{\operatorname{Trace}(A)} = \diagsum{3} + \diagsum{4} = 7 and multiply to det⁡(A)=(3)(4)−(1)(2)=10\detval{\det(A)} = (3)(4) - (1)(2) = 10. Because the eigenvalues multiply to det⁡(A)\detval{\det(A)}, having λ=0\eigval{\lambda} = \eigval{0} as an eigenvalue is the exact same thing as det⁡(A)=0\detval{\det(A)} = 0 (AA is not invertible).

7

Practice

1

Find the characteristic polynomial and the eigenvalues of A=[5234]A = \begin{bmatrix} 5 & 2 \\ 3 & 4 \end{bmatrix}.

Which entries of AA have λ\eigval{\lambda} subtracted from them before you take the 2×22\times 2 determinant?

Subtract λ\eigval{\lambda} from the main diagonal and set the determinant to zero:
det⁡[5−λ234−λ]=(5−λ)(4−λ)−(2)(3)=0\displaystyle \det\begin{bmatrix} 5 - \eigval{\lambda} & 2 \\ 3 & 4 - \eigval{\lambda} \end{bmatrix} = (5 - \eigval{\lambda})(4 - \eigval{\lambda}) - (2)(3) = 0
(Watch out: don't forget to subtract (2)(3)=6(2)(3) = 6. Leaving it off gives (5−λ)(4−λ)=λ2−9λ+20(5 - \eigval{\lambda})(4 - \eigval{\lambda}) = \eigval{\lambda}^2 - 9\eigval{\lambda} + 20 and the wrong eigenvalues λ=5,4\eigval{\lambda} = 5, 4.)
Expand and factor:
λ2−9λ+14\displaystyle \eigval{\lambda}^2 - 9\eigval{\lambda} + 14 =(λ−7)(λ−2)\displaystyle = (\eigval{\lambda} - 7)(\eigval{\lambda} - 2) =0\displaystyle = 0
Characteristic polynomial: λ2−9λ+14\lambda^2 - 9\lambda + 14. Eigenvalues: λ=7\lambda = 7 and λ=2\lambda = 2.
2

The matrix A=[1423]A = \begin{bmatrix} 1 & 4 \\ 2 & 3 \end{bmatrix} has eigenvalue λ=5\eigval{\lambda} = \eigval{5}. Use (A−5I)v=[00](A - \eigval{5}I)\eigvec{v} = \begin{bmatrix} 0 \\ 0 \end{bmatrix} to find a non-zero eigenvector v\eigvec{v} for λ=5\eigval{\lambda} = \eigval{5}, and verify that Av=5vA\eigvec{v} = \eigval{5}\eigvec{v}.

After subtracting 5\eigval{5} from the main diagonal, write out the equation from the top row for [xy]\eigvec{\begin{bmatrix} x \\ y \end{bmatrix}} and choose a simple non-zero number for x\eigvec{x}.

Subtract 5\eigval{5} from the main diagonal of AA:
A−5I=[1−5423−5]=[−442−2]\displaystyle A - \eigval{5}I = \begin{bmatrix} 1 - \eigval{5} & 4 \\ 2 & 3 - \eigval{5} \end{bmatrix} = \begin{bmatrix} -4 & 4 \\ 2 & -2 \end{bmatrix}
(Watch out: subtract 5\eigval{5} only from the main diagonal, not all four entries; and when solving −4x+4y=0-4\eigvec{x} + 4\eigvec{y} = 0, keep the sign straight so you don't write y=−xy = -x and get [1−1]\begin{bmatrix} 1 \\ -1 \end{bmatrix}, which gives A[1−1]=[−3−1]A\begin{bmatrix} 1 \\ -1 \end{bmatrix} = \begin{bmatrix} -3 \\ -1 \end{bmatrix}.)
Set (A−5I)[xy]=[00](A - \eigval{5}I)\eigvec{\begin{bmatrix} x \\ y \end{bmatrix}} = \begin{bmatrix} 0 \\ 0 \end{bmatrix}; the top row gives −4x+4y=0-4\eigvec{x} + 4\eigvec{y} = 0, so y=x\eigvec{y} = \eigvec{x}. Choosing x=1\eigvec{x} = 1 gives v=[11]\eigvec{v} = \eigvec{\begin{bmatrix} 1 \\ 1 \end{bmatrix}} (or any non-zero multiple).
Check: A[11]=[(1)(1)+(4)(1)(2)(1)+(3)(1)]A\eigvec{\begin{bmatrix} 1 \\ 1 \end{bmatrix}} = \begin{bmatrix} (1)(\eigvec{1}) + (4)(\eigvec{1}) \\ (2)(\eigvec{1}) + (3)(\eigvec{1}) \end{bmatrix} =[55]=5[11]{}= \begin{bmatrix} 5 \\ 5 \end{bmatrix} = \eigval{5}\eigvec{\begin{bmatrix} 1 \\ 1 \end{bmatrix}}.
3

Does the matrix A=[1−221]A = \begin{bmatrix} 1 & -2 \\ 2 & 1 \end{bmatrix} have any real eigenvalues? Use the characteristic equation to explain why or why not.

Write out det⁡(A−λI)=0\det(A - \eigval{\lambda} I) = 0 and pay close attention to the sign when you subtract (−2)(2)(-2)(2).

Set the determinant of A−λIA - \eigval{\lambda} I to zero:
det⁡[1−λ−221−λ]=(1−λ)2−(−2)(2)=λ2−2λ+5=0\displaystyle \det\begin{bmatrix} 1 - \eigval{\lambda} & -2 \\ 2 & 1 - \eigval{\lambda} \end{bmatrix} = (1 - \eigval{\lambda})^2 - (-2)(2) = \eigval{\lambda}^2 - 2\eigval{\lambda} + 5 = 0
(Watch out: dropping a minus sign in −(−2)(2)=+4-(-2)(2) = +4 gives (1−λ)2−4=λ2−2λ−3=0(1 - \eigval{\lambda})^2 - 4 = \eigval{\lambda}^2 - 2\eigval{\lambda} - 3 = 0, which leads to the wrong answer "Yes, λ=3\eigval{\lambda} = 3 and λ=−1\eigval{\lambda} = -1.")
Because (1−λ)2+4=0(1 - \eigval{\lambda})^2 + 4 = 0 has no real solutions (the matrix rotates every vector off its line), AA has no real eigenvalues.
4

Find the eigenvalues of the lower triangular matrix A=[400−3−10725]A = \begin{bmatrix} 4 & 0 & 0 \\ -3 & -1 & 0 \\ 7 & 2 & 5 \end{bmatrix}.

Look at the entries above the main diagonal. Do you need to expand a cubic determinant here?

Because every entry above the main diagonal is 00, the matrix is triangular and det⁡(A−λI)\det(A - \eigval{\lambda} I) =(4−λ)(−1−λ)(5−λ)=0{}= (4 - \eigval{\lambda})(-1 - \eigval{\lambda})(5 - \eigval{\lambda}) = 0.
(Watch out: setting (4−λ)=0(4 - \eigval{\lambda}) = 0 gives λ=4\eigval{\lambda} = \eigval{4}, not −4-4. Don't flip the signs of the diagonal entries and write λ=−4,1,−5\eigval{\lambda} = -4, 1, -5.)
λ=4\lambda = 4, λ=−1\lambda = -1, and λ=5\lambda = 5.
5

Suppose λ=0\eigval{\lambda} = \eigval{0} is an eigenvalue of A=[634k]A = \begin{bmatrix} 6 & 3 \\ 4 & k \end{bmatrix}. Find kk and the other eigenvalue of AA.

What does having λ=0\eigval{\lambda} = \eigval{0} as an eigenvalue tell you about det⁡(A)\detval{\det(A)}, and how does Trace⁡(A)\diagsum{\operatorname{Trace}(A)} give the second eigenvalue once kk is known?

Since λ=0\eigval{\lambda} = \eigval{0} is an eigenvalue, det⁡(A)=0\detval{\det(A)} = 0:
det⁡(A)\displaystyle \detval{\det(A)} =6k−(3)(4)\displaystyle = 6k - (3)(4) =6k−12\displaystyle = 6k - 12 =0⟶k\displaystyle = 0 \quad\longrightarrow\quad k =2\displaystyle = 2
(Watch out: λ=0\eigval{\lambda} = \eigval{0} makes the product det⁡(A)=0\detval{\det(A)} = 0, not the trace. Setting Trace⁡(A)=6+k=0\diagsum{\operatorname{Trace}(A)} = 6 + k = 0 gives the wrong value k=−6k = -6.)
Now use the trace rule λ1+λ2=Trace⁡(A)\eigval{\lambda_1} + \eigval{\lambda_2} = \diagsum{\operatorname{Trace}(A)} with k=2k = 2:
0+λ2\displaystyle \eigval{0} + \eigval{\lambda_2} =6+2\displaystyle = \diagsum{6} + \diagsum{2} =8\displaystyle = 8
k=2k = 2 and λ=8\lambda = 8.
6

Find all three eigenvalues of A=[010−1415−13]A = \begin{bmatrix} 0 & 1 & 0 \\ -1 & 4 & 1 \\ 5 & -1 & 3 \end{bmatrix} by expanding det⁡(A−λI)=0\det(A - \eigval{\lambda} I) = 0 and checking your result with the trace.

Expand along the top row using the +,−,++, -, + sign pattern, then check whether λ=1\eigval{\lambda} = \eigval{1} makes the resulting cubic equal to zero.

Subtract λ\eigval{\lambda} from the main diagonal and expand along the top row:
det⁡[−λ10−14−λ15−13−λ]=(−λ)((4−λ)(3−λ)+1)−1(−(3−λ)−5)=0\displaystyle \begin{aligned} & \det\begin{bmatrix} -\eigval{\lambda} & 1 & 0 \\ -1 & 4 - \eigval{\lambda} & 1 \\ 5 & -1 & 3 - \eigval{\lambda} \end{bmatrix} \\ &\quad = (-\eigval{\lambda})\big((4 - \eigval{\lambda})(3 - \eigval{\lambda}) + 1\big) \\ &\quad \qquad - 1\big(-(3 - \eigval{\lambda}) - 5\big) = 0 \end{aligned}
(Watch out: the second entry in the top row takes a minus sign in the +,−,++, -, + expansion; using +1+1 instead of −1-1 turns −(λ−8)-(\eigval{\lambda} - 8) into +(λ−8)+(\eigval{\lambda} - 8) and gives the wrong cubic λ3−7λ2+12λ+8=0\eigval{\lambda}^3 - 7\eigval{\lambda}^2 + 12\eigval{\lambda} + 8 = 0.)
Simplify and multiply by −1-1:
(−λ)(λ2−7λ+13)−1(λ−8)=0−λ3+7λ2−14λ+8=0λ3−7λ2+14λ−8=0\displaystyle \begin{aligned} (-\eigval{\lambda})(\eigval{\lambda}^2 - 7\eigval{\lambda} + 13) - 1(\eigval{\lambda} - 8) &= 0 \\ -\eigval{\lambda}^3 + 7\eigval{\lambda}^2 - 14\eigval{\lambda} + 8 &= 0 \\ \eigval{\lambda}^3 - 7\eigval{\lambda}^2 + 14\eigval{\lambda} - 8 &= 0 \end{aligned}
Testing λ=1\eigval{\lambda} = \eigval{1} gives 1−7+14−8=01 - 7 + 14 - 8 = 0, so (λ−1)(\eigval{\lambda} - 1) is a factor. Factoring out (λ−1)(\eigval{\lambda} - 1) gives:
(λ−1)(λ2−6λ+8)\displaystyle (\eigval{\lambda} - 1)(\eigval{\lambda}^2 - 6\eigval{\lambda} + 8) =(λ−1)(λ−2)(λ−4)\displaystyle = (\eigval{\lambda} - 1)(\eigval{\lambda} - 2)(\eigval{\lambda} - 4) =0\displaystyle = 0
Check with the trace: 1+2+4=7=0+4+3\eigval{1} + \eigval{2} + \eigval{4} = 7 = \diagsum{0} + \diagsum{4} + \diagsum{3} =Trace⁡(A){}= \diagsum{\operatorname{Trace}(A)}.
λ=1\lambda = 1, λ=2\lambda = 2, and λ=4\lambda = 4.
7

Let AA be a 3×33\times 3 matrix with Trace⁡(A)=9\diagsum{\operatorname{Trace}(A)} = 9, and suppose two of its eigenvalues are λ1=2\eigval{\lambda_1} = \eigval{2} and λ2=−1\eigval{\lambda_2} = \eigval{-1}. If B=P−1APB = P^{-1}AP for an invertible matrix PP, find all three eigenvalues of BB and det⁡(B)\detval{\det(B)}.

Use Trace⁡(A)\diagsum{\operatorname{Trace}(A)} to find the third eigenvalue of AA first; how do the eigenvalues and determinant of B=P−1APB = P^{-1}AP compare with those of AA?

The three eigenvalues of AA add up to Trace⁡(A)=9\diagsum{\operatorname{Trace}(A)} = 9:
2+(−1)+λ3\displaystyle \eigval{2} + (\eigval{-1}) + \eigval{\lambda_3} =9⟶λ3\displaystyle = 9 \quad\longrightarrow\quad \eigval{\lambda_3} =8\displaystyle = \eigval{8}
(Watch out: don't drop the negative sign on λ2=−1\eigval{\lambda_2} = \eigval{-1} and write 2+1+λ3=92 + 1 + \eigval{\lambda_3} = 9, which gives λ3=6\eigval{\lambda_3} = 6 and the wrong determinant det⁡(B)=(2)(−1)(6)=−12\detval{\det(B)} = (2)(-1)(6) = -12.)
Because B=P−1APB = P^{-1}AP is similar to AA, BB has the exact same eigenvalues as AA, and det⁡(B)\detval{\det(B)} is their product:
Eigenvalues of BB: λ=2,−1,8\lambda = 2, -1, 8, and det⁡(B)=(2)(−1)(8)=−16\det(B) = (2)(-1)(8) = -16.