When a matrix transforms space, it knocks almost every vector onto a new line, so we want to find the special directions that stay on their own line.
1
Directions that do not turn
An eigenvector of a matrix is a non-zero vector that stays on its own line when multiplied by the matrix, only getting stretched, shrunk, or flipped. Remember that multiplying a matrix by (x,y) takes x times the first column plus y times the second column.
Directions that do not turn
When a non-zero vector v stays on its line after multiplying by A, the output Av is just a number λ times the original vector v:
Aeigenvectorv=eigenvalueλv(v=0)
Here v cannot be the zero vector (0,0) because the zero vector has no direction, and a negative λ still counts because flipping a vector backwards keeps it on the exact same line.
Text solution
Test (1,1) by combining columnsTaking 1 of the first column plus 1 of the second column of A=[3012] sends (1,1) to (4,2), which turns off the line of (1,1).
[3012][11]=1[30]+1[12]=[42]
Check the horizontal vector (1,0)The first column shows that (1,0) lands at (3,0), staying on the horizontal axis and stretching by 3.
[3012][10]=[30]=3[10]
Test (−1,1) and factor out the stretchCombining −1 of the first column and 1 of the second column gives (−2,2), which is 2 times (−1,1) along the same diagonal line.
[3012][−11]=−1[30]+1[12]=[−22]=2[−11]
So any non-zero vector on the horizontal axis is an eigenvector with λ=3, and any non-zero vector on the line through (−1,1) is an eigenvector with λ=2.
For A=[3012], the transformed vector Av gets knocked off the line of v in almost every direction, staying on the same line only along the horizontal axis (where Av=3v) and the line through (−1,1) (where Av=2v).
This matrix knocks a vector off its line in every direction except along two lines, where it stays on the same line and only stretches.
2
Spotting eigenvectors from geometry
When we already know how a transformation moves the plane, such as a reflection, a shear, or a rotation, we can find its eigenvector lines and eigenvalues directly from the picture without doing any arithmetic.
Eigenvectors of reflections, shears, and rotations
Reflection
Horizontal shear
Rotation
Text solution
Find the unchanged and flipped lines of a reflectionReflecting across a line leaves vectors on the line of reflection fixed (λ=1) and flips vectors perpendicular to that line straight across into the opposite direction (λ=−1).
AvAv=1v(on the line),=−1v(perpendicular to the line)
Check the fixed axis of a horizontal shearA horizontal shear keeps vectors on the horizontal axis fixed (λ=1) and pushes every vector pointing up or down sideways off its line, leaving only one eigenvector line.
Av=1v(on the horizontal axis)
Check whether a rotation leaves any line in placeRotating the plane by an angle like 30∘ turns every vector onto a new line and gives no real eigenvectors, whereas a 180∘ rotation flips every non-zero vector in the plane backwards along its own line.
Av=−1v(every v=0 for a 180∘ rotation)
So a 2×2 transformation in the plane can have two eigenvector lines, only one, every line, or none at all.
3
What an eigenvalue of zero means
We have seen eigenvalues that stretch (λ=3), hold still (λ=1), and flip (λ=−1). Now let's see what happens when λ=0. Remember that the vectors a matrix sends to (0,0) form its null space, and a matrix sends non-zero vectors to (0,0) only when it flattens space onto a lower dimension.
What an eigenvalue of zero means
Setting λ=0 in Av=λv turns the eigenvector equation directly into the null space equation:
Av=0v=0
So the eigenvectors for λ=0 are the non-zero vectors in the null space of A, and saying λ=0 is an eigenvalue of A means the exact same thing as saying det(A)=0 and A is not invertible.
4
Finding 2×2 eigenvectors from A−λI
For a matrix like A=[4213], both basis vectors turn, so we cannot spot the eigenvectors from the picture. Once we are given an eigenvalue λ, we can find its eigenvectors by rewriting "multiply by λ" as the matrix λI=[λ00λ].
Finding 2×2 eigenvectors from A − λI
Rewriting λv as λIv and subtracting it from Av subtracts λ from the diagonal of A and turns every eigenvector problem into a null space problem:
(A−λI)v=0
For A−λI to have non-zero vectors in its null space, it must flatten space (det(A−λI)=0). For a 2×2 matrix, that means the two columns of A−λI always land on the same line whenever λ is an eigenvalue. (If the two columns of a 2×2 matrix A−λI are not multiples of each other, stop and check your subtraction: either there is an arithmetic mistake or λ is not an eigenvalue.)
Text solution
Subtract λ=5 from the diagonalSubtract 5 from the diagonal entries of A=[4213] to get A−5I.
A−5I=[4−5213−5]=[−121−2]
Cancel the two columns for λ=5The two columns (−1,2) and (1,−2) point in opposite directions on the same line, so 1 of the first column plus 1 of the second column cancels to (0,0).
1[−12]+1[1−2]=[00]⟹v=c[11](c=0)
Subtract λ=2 from the diagonalSubtract 2 from the diagonal entries of A to get A−2I.
A−2I=[4−2213−2]=[2211]
Cancel the two columns for λ=2The first column (2,2) is twice the second column (1,1), so 1 of the first column minus 2 of the second column cancels to (0,0).
1[22]−2[11]=[00]⟹v=c[1−2](c=0)
So A stretches every non-zero vector on the line c(1,1) by 5 and every non-zero vector on the line c(1,−2) by 2.
Because the two columns (−1,2) and (1,−2) of A−5I lie on the same line, any vector v=(x,y) on the line through (1,1) makes the column combination x(−1,2)+y(1,−2) cancel to (0,0), which is the exact line where Av equals 5v.
Every vector on the diagonal line makes the two opposite columns on the left cancel out, which is the exact line where the matrix on the right stretches the vector by 5 without turning it.
5
Testing eigenvalues and finding 3×3 eigenvectors by elimination
In 2D, you can immediately see whether two columns line up. For a 3×3 matrix in 3D space, you usually cannot tell just by looking whether three columns flatten into a plane or what combination cancels them out. When we are given a number λ to test for a 3×3 matrix, we want to check whether λ is an eigenvalue and find its eigenvectors at the same time. Remember that row elimination finds the null space of a matrix: if every column gets a pivot, no variables are free and only (0,0,0) is in the null space; if a column has no pivot, its variable is free and gives us non-zero solutions.
Testing eigenvalues and finding 3×3 eigenvectors by elimination
Aλv=311254137=4?λ=2?=?
Text solution
Eliminate A−4I down the columnsSubtract 4 from the diagonal of A=311254137 and use row operations to clear the entries below the diagonal.
Set the free variable z=1 and solveThe row of zeros leaves the third column without a pivot, so z is free, λ=2 is an eigenvalue, and setting z=1 in the two remaining equations gives the eigenvectors.
x+2y+zy+2z=0=0}z=1y=−2,x=3⟹v=c3−21(c=0)
So λ=4 is not an eigenvalue of A (no direction in 3D space stretches by 4), whereas λ=2 is an eigenvalue whose eigenvectors c(3,−2,1) stay on their line and double in length.
6
Applying a matrix many times with eigenvectors
Now suppose we want to multiply a starting vector like (3,0) by A=[4213] a hundred times in a row. Doing that by hand is a mess because the x and y coordinates of the vector mix at every step, so we break the starting vector into pieces along the eigenvector lines instead.
Applying a matrix many times with eigenvectors
AA100[30]=[4213]=?
Text solution
Split the starting vector along the eigenvectorsWrite (3,0) as a combination of the eigenvectors (1,1) (which stretches by λ=5) and (1,−2) (which stretches by λ=2).
[30]=2[11]+1[1−2]
Multiply each eigenvector piece by λ100Because A stretches each eigenvector along its own line without turning or mixing the two pieces, multiplying by A a hundred times scales the first piece by 5100 and the second piece by 2100.
A100[30]=2⋅5100[11]+1⋅2100[1−2]
Because 5100 is much larger than 2100, multiplying by A over and over pulls the vector closer and closer to the line of the (1,1) eigenvector.
When any starting vector v off the (1,−2) line is split into pieces along the eigenvector lines through (1,1) and (1,−2), each multiplication by A stretches the (1,1) piece by 5 and the (1,−2) piece by 2 without mixing them, pulling Av,A2v, and A3v toward the faster-growing (1,1) line.
Each step stretches the two pieces along their own lines at different speeds, pulling the vector toward the faster-growing line whenever it has a piece on that line.
Key concepts
An eigenvector v=0 is a direction that does not turn when multiplied by A, so Av=λv. Before setting up A−λI, check whether you already have the vector or the geometry in front of you:
Testing a given vector v (do not use A−λI)
If you are given a specific vector v, just multiply Av and check if the result is a multiple of v. For A=[3012] and v=(−1,1),Av=(−2,2)=2(−1,1), so v is an eigenvector with λ=2.
Known geometry in the plane (do not compute)
Look for directions that stay fixed (λ=1, such as a line of reflection or a shear's fixed axis) and directions that flip straight backwards (λ=−1, such as the line perpendicular to a reflection or every vector in a 180∘ rotation). A rotation that turns every line (like 30∘ or 90∘) has no real eigenvectors.
Collapsed columns (λ=0)
If the columns of A already lie on the same line (det(A)=0, not invertible), then λ=0 is an eigenvalue and its eigenvectors are the non-zero vectors in the null space of A (Av=0).
Finding eigenvectors from a given λ
Subtract λ from the diagonal to turn Av=λv into the null space equation
(A−λI)v=0
For a 2×2 matrix like A−2I=[2211], look for the weights that cancel the two columns: 1[22]−2[11]=[00] gives v=c(1,−2). For a 3×3 matrix (or to test whether a number λ is an eigenvalue), run row elimination on A−λI: a free variable (a column with no pivot) confirms λ is an eigenvalue and gives the eigenvectors, while a pivot in every column forces v=(0,0,0) and rules λ out.
Multiplying by A many times (Akv)
Never multiply A by itself. Split the starting vector into eigenvector pieces and multiply each piece by λk: for (3,0)=2(1,1)+1(1,−2) with λ1=5 and λ2=2,A100(3,0)=2⋅5100(1,1)+1⋅2100(1,−2).
7
Practice
1
Let A=[1423]. Test each vector u=(1,1),v=(1,2), and w=(1,−1) by multiplying by A to see whether it is an eigenvector of A, and if so, give its eigenvalue λ.
Watch out: do not look only at the first coordinate of (3,7) and call λ=3, because 3(1,1)=(3,3)=(3,7).
Multiply A by each vector by combining the columns of A: is the resulting vector a single number λ times the starting vector?
Multiply each vector by A=[1423]:
Au=1[14]+1[23]=[37]
Watch out: do not look only at the first coordinate of (3,7) and call λ=3, because 3(1,1)=(3,3)=(3,7). Since (3,7) turns off the line of (1,1),u is not an eigenvector.
Av=1[14]+2[23]=[510]=5[12]
Aw=1[14]−1[23]=[−11]=−1[1−1]
Be careful on w:(−1,1) is the exact opposite of (1,−1) along the same line, so don't throw it out or write λ=1. Factoring out −1 gives λ=−1.
u=(1,1) is not an eigenvector; v=(1,2) is an eigenvector with λ=5;w=(1,−1) is an eigenvector with λ=−1.
2
Without doing any calculations, find all real eigenvector lines and their eigenvalues in the 2D plane for: (a) reflection across the line y=x, (b) a vertical shear that keeps the vertical axis fixed, and (c) a 90∘ counter-clockwise rotation versus a 180∘ rotation.
Picture which lines of vectors stay completely still (λ=1) and which lines flip straight across into the opposite direction (λ=−1).
(a) For reflection across y=x, non-zero vectors on the line y=x stay fixed with λ=1. Watch out: a common mistake is stopping after the reflection line and forgetting the perpendicular line y=−x, where every vector flips straight across along its own line with λ=−1. (b) For a vertical shear that keeps the vertical axis fixed, only vectors on the vertical axis stay on their line (don't pick the horizontal axis c(1,0) out of habit from a horizontal shear, because a vertical shear pushes (1,0) up or down off its line). (c) A 90∘ rotation turns every vector onto a perpendicular line and has no real eigenvectors. Don't assume a 180∘ rotation also has none just because it is a rotation: turning by 180∘ flips every vector in the plane straight backwards along its own line with λ=−1.
(a) c(1,1) with λ=1 and c(1,−1) with λ=−1 (c=0); (b) c(0,1) with λ=1 (c=0); (c) 90∘ rotation: no real eigenvectors; 180∘ rotation: every non-zero vector in the plane with λ=−1.
3
Let A=[2−1−63]. By looking at its columns, explain why λ=0 is an eigenvalue of A, state whether A is invertible, and find all eigenvectors for λ=0.
Watch out: because the second column is −3 times the first, it is easy to write c(−3,1) or c(1,−3) by mistake, which gives (−12,6) or (20,−10) instead of (0,0); you need the weights 3 and 1 that make the two columns cancel.
Compare the two columns of A: do they point in different directions, or is one a multiple of the other?
The second column (−6,3) is −3 times the first column (2,−1), so the columns collapse onto one line, det(A)=0, and A is not invertible. Because A crushes non-zero vectors to (0,0),λ=0 is an eigenvalue whose eigenvectors are the non-zero vectors in the null space of A:
3[2−1]+1[−63]=[00]
Watch out: because the second column is −3 times the first, it is easy to write c(−3,1) or c(1,−3) by mistake, which gives (−12,6) or (20,−10) instead of (0,0); you need the weights 3 and 1 that make the two columns cancel.
λ=0 is an eigenvalue because the columns are multiples of each other, A is not invertible, and the eigenvectors are c(3,1) for c=0.
4
The matrix A=[5324] has eigenvalues λ=7 and λ=2. Use A−λI to find the eigenvectors for each eigenvalue.
Subtract λ only from the diagonal entries of A, then find the weights that make the two columns of A−λI cancel to (0,0).
Watch out on λ=2: don't just copy a row of A−2I and write c(3,2) (which gives 3(3,3)+2(2,2)=(13,13)=(0,0)). You need 2 of the first column and −3 of the second column to make them cancel, which gives c(2,−3).
For λ=7:c(1,1) (c=0); for λ=2:c(2,−3) (c=0).
5
Let A=411153237. Use row elimination on A−λI to test whether λ=5 and λ=3 are eigenvalues of A, and find the eigenvectors for any that is.
Be careful with signs when solving for x and y: forgetting to flip the sign when moving terms across the equals sign gives the common wrong answer c(3,1,1) or c(1,−1,1) instead of c(−1,−1,1).
Subtract λ from all three diagonal entries and eliminate down the columns; look for whether all three columns get pivots or if a row of zeros leaves a free variable.
First test λ=5 by subtracting 5 from the diagonal and eliminating:
Watch out: don't get fooled by the 0 in the middle of A−5I before eliminating. Once you finish elimination, every column has a pivot (−1,1, and −16) and no variable is free, so the null space is only (0,0,0) and λ=5 is not an eigenvalue.
Now test λ=3 by subtracting 3 from the diagonal and eliminating:
The third column has no pivot, so z is free and λ=3 is an eigenvalue. Setting z=1 in the two remaining equations x+y+2z=0 and y+z=0:
y+1=0x+(−1)+2(1)=0⟹y=−1,⟹x=−1
Be careful with signs when solving for x and y: forgetting to flip the sign when moving terms across the equals sign gives the common wrong answer c(3,1,1) or c(1,−1,1) instead of c(−1,−1,1).
λ=5 is not an eigenvalue; λ=3 is an eigenvalue with eigenvectors c(−1,−1,1) for c=0.
6
Using the eigenvectors v1=(1,1) (with λ=7) and v2=(2,−3) (with λ=2) of A=[5324], find the result of multiplying the starting vector (5,0) by A50 times.
Watch out: do not raise the weights 3 and 1 to the 50th power (which gives the wrong answer 350⋅750(1,1)+250(2,−3)). The weights 3 and 1 are found once at the start, and each multiplication by A multiplies the eigenvector pieces by λ.
First find how many of v1=(1,1) and how many of v2=(2,−3) combine to make (5,0).
Write the starting vector (5,0) as a combination of v1=(1,1) and v2=(2,−3):
[50]=3[11]+1[2−3]
Multiplying by A50 times stretches the (1,1) piece by 750 and the (2,−3) piece by 250:
A50[50]=3⋅750[11]+1⋅250[2−3]
Watch out: do not raise the weights 3 and 1 to the 50th power (which gives the wrong answer 350⋅750(1,1)+250(2,−3)). The weights 3 and 1 are found once at the start, and each multiplication by A multiplies the eigenvector pieces by λ.