Multiplying a vector by a matrix, or multiplying a matrix by itself many times, usually mixes the coordinates together, so we want a coordinate grid where the matrix only stretches.
1
Stretching Along the Eigenbasis
We want to see why a matrix like [3113], which mixes x and y at every step, is geometrically doing the same stretches as the diagonal matrix [4002], just along tilted axes.
Stretching Along the Eigenbasis
In standard x-y coordinates, [3113] mixes both entries, while along its two eigenvector directions it stretches without any mixing at all:
[3113][11][3113][1−1]=4[11],=2[1−1]
2
The Three-Step Detour Through the Eigenbasis
We want to compute [3113][42] by switching into the eigenbasis v1=[11] and v2=[1−1], where the transformation is a simple diagonal stretch.
The Three-Step Detour Through the Eigenbasis
[3113][42]=?
Text solution
1. Switch from ordinary coordinates to eigenbasis coordinatesWrite [42] as a combination of the eigenvectors v1=[11] and v2=[1−1] to see how much of each eigenvector it takes.
2. Stretch each eigenbasis coordinate by its eigenvalueMultiply by the diagonal matrix of eigenvalues [4002] so the first coordinate stretches by 4 and the second stretches by 2.
[4002][31]=[122]
3. Translate back to ordinary coordinatesCombine the eigenvectors using the stretched eigenbasis coordinates as weights to get the final vector in standard x-y coordinates.
12[11]+2[1−1]=[12+212−2]=[1410]
Pick any vector x in the plane: [3113] mixes its standard x and y coordinates, while its two components along the eigenvector lines [11] and [1−1] simply stretch by 4 and 2 to land on [3113]x.
While this matrix mixes standard x and y coordinates together, on the eigenvector grid it simply stretches one tilted component by 4 and the other by 2.
3
Diagonalization: A=PDP−1 and Matrix Powers
Diagonalization writes the three-step detour through the eigenbasis as a single matrix product A=PDP−1, which lets us raise a matrix to any power Ak without mixing coordinates at every step.
Remember that putting the new basis vectors into the columns of a matrix P translates from new basis coordinates into ordinary coordinates, while its inverse P−1 translates the other way, from ordinary coordinates into the new basis.
Diagonalization: A = PDP⁻¹ and Matrix Powers
When PDP−1 multiplies a vector x, the three matrices act in order from right to left:
1. Put the eigenvectors into P and the eigenvalues into DPut the eigenvectors into the columns of P and place each matching eigenvalue into the same column of the diagonal matrix D.
P=[111−1],D=[4002]
2. Invert P to complete A=PDP−1Find P−1 for the step that translates into the eigenbasis, giving A=PDP−1.
3. Cancel matching P−1P pairs to find AkWrite out PDP−1 multiplied by itself and cancel each inner P−1P pair to the identity matrix, so only the diagonal matrix D is raised to the k-th power.
Diagonalizing a 3×3 Matrix with a Repeated Eigenvalue
Now we want to diagonalize the 3×3 matrix A=301121103 as A=PDP−1, where a repeated eigenvalue has to give us two linearly independent eigenvectors to fill the three columns of P.
Diagonalizing a 3×3 Matrix with a Repeated Eigenvalue
AP=301121103=0?00?00?0
Text solution
1. Find the eigenvalues from det(A−λI)=0Expand the determinant along the middle row and factor to find the roots and how many times each root repeats.
2. Find two independent eigenvectors from the repeated eigenvalue λ=2Row-reduce A−2I to get the plane equation x+y+z=0, then set the two free variables (y,z) to (1,0) and (0,1) in turn.
4. Put together P and D in matching column order and compute P−1Place v1,v2,v3 into the columns of P, put their matching eigenvalues 2,2,4 on the diagonal of D, and invert P.
You can put the eigenvectors into the columns of P in any order you like, as long as column i of D has the eigenvalue that goes with column i of P.
Every vector in the eigenspace plane x+y+z=0 stretches by 2 along its own line. So long as we pick v2 in that plane, not on the same line as v1=−110, it teams up with v1 and v3=101 to make a full 3D eigenbasis grid where PDP−1=A. If we line v2 up with v1, the 3D grid collapses into a plane and P has no inverse.
Every direction on the eigenspace plane stretches by two, and any second vector on that plane not lined up with the first completes a full 3D eigenbasis to diagonalize the matrix.
5
When Is a Matrix Diagonalizable?
An n×n matrix is diagonalizable when it has n linearly independent eigenvectors to form a complete coordinate grid for n-dimensional space and make P invertible.
When Is a Matrix Diagonalizable?
AP=PDP−1=[v1v2v3]for every matrix?
If an n×n matrix has n different real eigenvalues, their eigenvectors automatically point in linearly independent directions, so the matrix is guaranteed to be diagonalizable without checking any eigenspaces. When an eigenvalue repeats, we check whether the matrix B=200120004 still has enough independent eigenvectors:
Text solution
1. Read the eigenvalues and their algebraic multiplicities from the diagonalBecause B is upper-triangular, its eigenvalues are the diagonal entries.
3. Compare geometric multiplicity with algebraic multiplicityBecause λ=2 has algebraic multiplicity 2 but only gives a 1-dimensional line of eigenvectors, B has only two independent eigenvectors in total and cannot fill the three columns of P.
geometric multiplicity (1)<algebraic multiplicity (2)⟹B is not diagonalizable
Under the 2D shear matrix [1011] with double eigenvalue λ=1, every unit vector v off the horizontal axis is pushed sideways off its own line, leaving only the line along [10] on its own line and no second independent eigenvector direction to build a 2D coordinate grid.
Even though its eigenvalue repeats twice, this shear pushes every vector off the horizontal axis away from its own line, leaving only one eigenvector direction instead of the two needed for a 2D grid.
Key concepts
Diagonalizing an n×n matrix A writes it as a three-step detour through its eigenbasis, A=PDP−1. The columns of P are n linearly independent eigenvectors, and the diagonal entries of D are the matching eigenvalues in the same column order (for A=[3113],P=[111−1] and D=[4002]). When A multiplies a vector x,P−1x translates ordinary coordinates into eigenbasis coordinates, DP−1x stretches each coordinate independently, and PDP−1x translates back to ordinary coordinates. When you raise A to a power, every P−1P in the middle cancels out, so Ak=PDkP−1 only raises the diagonal entries of D to the k-th power. And if a matrix is already diagonal, like [4002]10=[41000210], its coordinates already don't mix, so no basis change is needed.
To see if an n×n matrix is diagonalizable, start with the roots of det(A−λI)=0. If all n eigenvalues are real and different (like λ=1,4,−2 for a 3×3 matrix), you can stop right there: eigenvectors from different eigenvalues are automatically linearly independent, so the matrix is diagonalizable without solving for any eigenspaces. You only need to check A−λI when an eigenvalue repeats. Then the matrix is diagonalizable when every repeated eigenvalue has as many free variables in A−λI (its geometric multiplicity) as the number of times that root repeats (its algebraic multiplicity), like λ=2 giving two free variables in x+y+z=0 versus only one free variable in 000100002.
6
Practice
1
A 2×2 matrix A has eigenvectors v1=[12] with eigenvalue λ1=5 and v2=[11] with eigenvalue λ2=3. Use the three-step detour through the eigenbasis to compute Ax for x=[35]: (a) find the eigenbasis coordinates P−1x, (b) multiply by D to find DP−1x, and (c) translate back to ordinary coordinates to find Ax.
For (a), ask how many of v1=[12] and how many of v2=[11] add up to [35].
(a) Write x=[35] as a combination of v1 and v2:
[35]=2[12]+1[11]⟹P−1x=[21]
(b) Multiply the eigenbasis coordinates P−1x by D=[5003] (don't multiply the original coordinates [35] by D, which gives the wrong vector [1515]):
DP−1x=[5003][21]=[103]
(c) Use the stretched eigenbasis coordinates as weights on v1 and v2 to return to ordinary coordinates (don't stop at [103], which is still in eigenbasis coordinates):
Ax=10[12]+3[11]=[1323]
2
Diagonalize A=[64−20] as A=PDP−1, and use Ak=PDkP−1 to find A10 (leave powers of 4 and 2 in exponent form).
Raising the entries of A directly to the 10th power gives [610410(−2)100].
Find the eigenvalues from det(A−λI)=0, match each eigenvector column in P with its eigenvalue column in D, and multiply PD10P−1.
Solve the characteristic equation det(A−λI)=0:
(6−λ)(0−λ)−(−2)(4)=λ2−6λ+8=(λ−4)(λ−2)=0⟹λ=4,2
For λ=4,(A−4I)v=[24−2−4][xy]=[00] gives x=y, so v1=[11]. For λ=2,(A−2I)v=[44−2−2][xy]=[00] gives y=2x, so v2=[12]. Putting these into P and D and inverting P gives:
PDP−1=[1112],=[4002],=[2−1−11]
Compute A10=PD10P−1 with P on the left and P−1 on the right (swapping P and P−1 to P−1D10P gives the wrong answer [2⋅410−210−410+2102⋅410−2⋅210−410+2⋅210], and raising the entries of A directly to the 10th power gives [610410(−2)100]):
Determine whether each matrix is diagonalizable by comparing algebraic and geometric multiplicities: (a) A=40013025−1 and (b) B=300030123.
Both matrices are upper-triangular. Read the eigenvalues off the diagonal first and check whether any eigenvalue repeats.
(a) Reading the diagonal entries of A gives three different eigenvalues, λ=4,3,−1, so A automatically has 3 linearly independent eigenvectors: A is diagonalizable.
(b) Reading the diagonal entries of B gives λ=3 with algebraic multiplicity 3 (don't assume B is already diagonal just because all three diagonal entries are 3, since the non-zero entries in the last column still mix z into x and y). Checking B−3I:
B−3I=000000120⟹z=0(x and y free)
Because only x and y are free, the eigenspace has geometric multiplicity 2<3:B is not diagonalizable.
5
For what value(s) of k is M=200k20504 diagonalizable? For that value of k, find P and D such that M=PDP−1.
The repeated eigenvalue λ=2 has algebraic multiplicity 2, so look at M−2I and see what value of k leaves two free variables.
Because M is upper-triangular, its eigenvalues are λ=2 (algebraic multiplicity 2) and λ=4 (algebraic multiplicity 1). Subtracting 2I gives:
M−2I=000k00502
The third row forces 2z=0⟹z=0, which reduces the first row to ky=0. To get a 2D eigenspace with both x and y free, we need k=0.
When k=0, setting (x,y)=(1,0) and (0,1) with z=0 gives v1=100 and v2=010. For λ=4,M−4I=−2000−20500 gives y=0 and −2x+5z=0⟹2x=5z, so setting z=2 gives v3=502 (don't read −2x+5z=0 directly as x=−2,z=5 or x=2,z=5, which gives the wrong third column −205 or 205):