A linear transformation is a physical action on space, like a stretch or a rotation, and the matrix we write for it depends on the coordinate grid we lay down to measure it.
1
Same Map, Different Grids
Two matrices are similar when they describe the exact same linear map in two different coordinate bases.
Before any grid is drawn, stretching a flat sheet by a factor of 3 along the 45∘ line while leaving the perpendicular direction unchanged is a single geometric action. Laying down a coordinate grid only decides how we record where the basis vectors land.
Same Map, Different Grids
Recording that single 45∘ stretch in our standard grid versus a friend's tilted grid (b1 along the 45∘ stretch, b2 perpendicular to it) gives two different matrices for the same map:
Standard grid (^,^)A=[2112]andTilted grid (b1,b2)B=[3001]
As the coordinate grid turns over the stretched sheet, all four entries of the matrix change while the 45∘ oval stays fixed, and the off-diagonal entries drop to 0 right when the basis vectors line up with the stretch and its perpendicular direction.
Turning the coordinate grid changes the matrix for the exact same 45° stretch, and the off-diagonal numbers drop to 0 when the grid lines up with the stretch.
2
The Three-Step Detour: B=P−1AP
Suppose we know the matrix A=[2112] for a map in our standard grid, and we want to find the matrix B that describes that same map in the tilted basis b1=[11],b2=[−11].
To connect the two grids, we put the new basis vectors into the columns of a change-of-basis matrix P:
P=[b1b2]=[11−11]
Multiplying by P takes coordinates from their tilted grid and gives the vector in our standard grid, while P−1 translates from our standard grid back into their grid:
Their grid → Our gridP and Our grid → Their gridP−1
Because A only accepts standard coordinates, we cannot multiply A directly by a vector written in their grid. Instead, we translate the vector into our grid, apply A, and translate the answer back.
The Three-Step Detour: B = P⁻¹AP
b1
b2
A=[2112]⟷?B=[3001]
Reading from right to left in the order the matrices multiply an input vector x in their grid gives the formula for the similar matrix B:
B=3. Back totheir gridP−12. Map inour gridA1. To ourgridP
Text solution:
Transform the first basis vector in our grid and write it in their grid. We apply A to b1=[11] in standard coordinates and write the output as a combination of b1 and b2 to get Column 1 of B.
Ab1=[2112][11]=[33]=3b1+0b2⟹Column 1 of B=[30]
Transform the second basis vector in our grid and write it in their grid. We apply A to b2=[−11] in standard coordinates and write the output in terms of b1 and b2 to get Column 2 of B.
Ab2=[2112][−11]=[−11]=0b1+1b2⟹Column 2 of B=[01]
Put the new-basis columns together into B. Placing the two coordinate columns side by side gives the matrix for the stretch in the tilted basis.
B=[3001]
When the stretched vectors Ab1 and Ab2 are simple multiples of b1 and b2, you can see their new coordinates right away in Step 3 without writing out P−1.
For every input vector x in the tilted grid, translating to the standard grid with P, stretching with A, and translating back with P−1 lands on the exact same arrow and tilted coordinates as multiplying by B directly.
Translating any arrow to our standard grid, stretching it there, and translating it back always gives the exact same arrow and tilted coordinates as stretching it on the tilted grid directly.
3
Changing to a Skewed Basis
In the tilted basis above, the basis vectors lined up with the stretch, so we could see the new coordinates in Step 3 right away. Now we want the matrix B for the same stretch A=[2112] in a skewed basis u1=[23],u2=[12] that does not line up with the stretch at all.
When you cannot tell by looking how many u1's and u2's it takes to build the stretched vectors, multiplying on the left by P−1 does Step 3 for you.
Changing to a Skewed Basis
u1
u2
AB=[2112]=?
Text solution:
Put the skewed basis into P and find its inverse P−1. Placing u1 and u2 into the columns of P gives the translator to our grid, and P−1 translates back to the skewed grid.
PP−1=[2312],=[2−3−12]
Transform both basis vectors in the standard grid (AP). Multiplying A by P applies the map to u1 and u2 at once in standard coordinates.
AP=[2112][2312]=[7845]
Translate the outputs back into the skewed basis (P−1AP). Multiplying AP on the left by P−1 converts the standard vectors [78] and [45] into combinations of u1 and u2.
B=P−1(AP)=[2−3−12][7845]=[6−53−2]
So in the skewed basis u1,u2, the same 45∘ stretch is described by B=[6−53−2], where Column 1 tells us Au1=6u1−5u2 and Column 2 tells us Au2=3u1−2u2.
4
Back to the Standard Grid: A=PBP−1
Now suppose we start from the geometry of the stretch along b1=[11] and b2=[−11], so we already know the easy matrix B=[3001] in their tilted grid and want to find A in our standard grid.
To transform a vector that starts in our standard grid, we take the same three-step detour in reverse: translate from our grid to their tilted grid, apply the easy matrix B, and translate back to our grid.
Back to the Standard Grid: A = PBP⁻¹
b1
b2
^
^
BA=[3001]=?
Reading from right to left for an input vector starting in our standard grid gives A:
A=3. Back toour gridP2. Map intheir gridB1. To theirgridP−1
Text solution:
Write the tilted-basis matrix P and its inverse P−1. The columns of P hold b1 and b2, and P−1 translates standard vectors like ^ and ^ into tilted coordinates.
PP−1=[11−11],=[21−212121]
Apply the stretch in the tilted grid and translate back (PB). Multiplying P by B scales the basis columns by their stretch factors 3 and 1.
PB=[11−11][3001]=[33−11]
Translate from standard inputs on the right (PBP−1). Multiplying PB on the right by P−1 finishes the detour so A acts on standard coordinates.
A=(PB)P−1=[33−11][21−212121]=[2112]
This detour also gives a fast way to compute high powers of A: rather than multiplying the harder matrix A ten times in the standard grid, translate to the tilted basis once, apply the diagonal matrix B ten times, and translate back at the end:
A10=PB10P−1
5
Testing Whether Matrices Are Similar
Given two matrices, we want to find out whether they are similar, meaning they can represent the same map in two different bases, without solving for a basis matrix P.
Testing Whether Matrices Are Similar
A=[2112]C=[3112]
similar?
Because similar matrices represent the exact same geometric map, every property that belongs to the map itself must match across all of them:
Invertibility and rank (whether the map is reversible or flattens space)
Trace (the sum of the main diagonal entries):
trace of [2112]2+2=trace of [3001]3+1=trace of [6−53−2]6+(−2)=4
Determinant (how much areas scale) and eigenvalues (stretch factors)
I=[1001]
S=[1011]
^
^
^
^
[11]
Text solution:
Rule out similarity when a map property disagrees. Comparing A=[2112] and C=[3112], the sums of their main diagonals do not match, so they cannot represent the same map in any basis.
trace(A)=2+2=4=trace(C)=3+2=5⟹A and C are not similar
Check what P−1AP does when all the numbers match. The identity I=[1001] and the shear S=[1011] have the same trace (2), rank (2), determinant (1), and eigenvalues (1), but changing the basis on I always produces I, never S.
P−1IP=P−1P=I=[1011]⟹I and S are not similar
Across every skewed basis u1,u2 for the stretch A=[2112], the individual entries of B=P−1AP change a lot, but if one main diagonal entry goes up, the other goes down by the same amount so their sum stays at 4 (as long as u1 and u2 do not line up and flatten the grid).
Changing the basis can change all four numbers in the matrix, but in every working grid the main diagonal still adds up to 4.
Key concepts
Two matrices A (in the standard grid) and B (in a new basis whose vectors form the columns of P=[b1b2]) are similar when they describe the same linear map in two different coordinate systems. Whenever you convert between them, read the three-step detour from right to left based on which grid the input vector starts in:
Standard grid to a new basis (B=P−1AP)
An input vector starts in their grid, so P translates it to our grid first, A transforms it, and P−1 translates the result back. Before inverting P, check Ab1 and Ab2. If they are already obvious combinations of b1 and b2 (such as A[11]=[33]=3b1+0b2), write the columns of B directly without computing P−1. Compute P−1AP when the basis is skewed and you cannot see the combination right away.
New basis back to the standard grid (A=PBP−1)
When a map is easy to write in a tilted basis (like B=[3001]), an input vector in our standard grid is multiplied by P−1 first to enter their grid, gets transformed by B, and returns to our grid through P. Use this same detour An=PBnP−1 to compute powers of A by raising the diagonal matrix B to the nth power in the middle.
Testing similarity without solving for P
Check properties of the map itself first: trace (sum of the main diagonal), invertibility and rank, determinant, and eigenvalues (stretch factors). Any mismatch (like trace[2112]=4=trace[3112]=5) immediately rules out similarity. If all checks match, check how the map acts on vectors rather than assuming similarity: a matrix that scales every vector equally, like I, satisfies P−1IP=I in every basis and can never be similar to a shear like [1011].
6
Practice
1
In the standard grid, a linear map has the matrix A=[3212]. Find the similar matrix B=P−1AP that describes this map in the new basis b1=[11],b2=[1−2].
Before finding P−1, multiply Ab1 and Ab2 in the standard grid. How do those output vectors compare to b1 and b2?
Apply A to each basis vector in the standard grid and write the result in terms of b1 and b2:
Ab1=[3212][11]=[44]=4b1+0b2
Ab2=[3212][1−2]=[1−2]=0b1+1b2
(Watch out: don't stop at the standard outputs [44] and [1−2] and put them into [441−2]. The columns of B are the coordinates [40] and [01] in the new basis b1,b2.)
B=[4001]
2
Let A=[1123] be the matrix of a map in the standard basis. Write the change-of-basis matrix P and its inverse P−1 for the skewed basis u1=[11],u2=[12], and compute the similar matrix B=P−1AP in this basis.
Put u1 and u2 into the columns of P, then follow the three-step detour from right to left starting with an input in the skewed basis.
Place u1 and u2 into the columns of P and invert it:
P=[1112],P−1=[2−1−11]
Transform the basis columns in the standard grid first:
AP=[1123][1112]=[3457]
Translate the outputs back to the skewed basis by multiplying on the left by P−1(putting P on the left and P−1 on the right computes PAP−1=[−1−235], which translates in the wrong direction):
B=P−1AP=[2−1−11][3457]=[2132]
PP−1B=[1112],=[2−1−11],=[2132]
3
A transformation of the plane stretches vectors along b1=[11] by a factor of 3 and flips vectors along the perpendicular direction b2=[−11] to their negatives. Write the matrix B for this map in the basis b1,b2, and use A=PBP−1 to find its standard matrix A.
Since b2 is flipped to −b2, what is the second column of B in the b1,b2 grid? Then take the detour that starts in our standard grid.
In the basis b1,b2,b1 goes to 3b1+0b2 and b2 goes to 0b1−1b2(watch out: flipping b2 means multiplying by −1, not 0 or +1):
B=[300−1]
Write P and its inverse P−1(don't forget to divide by the determinant 1(1)−(−1)(1)=2 when finding P−1, or A will come out doubled as [2442]):
The matrix A=[31−20] satisfies A=PBP−1 for the basis P=[2111] (with P−1=[1−1−12]) and the diagonal matrix B=[2001]. Use the three-step detour to compute A5.
Translate to the tilted basis once, apply the diagonal stretch five times in a row, and translate back to the standard grid at the end.
Raise the diagonal matrix B to the 5th power in the tilted basis:
B5=[250015]=[32001]
Use A5=PB5P−1(be careful not to swap P and P−1 into P−1B5P, which gives [63−6231−30], and never raise the individual entries of A directly to the 5th power):
Consider the three matrices M=[2142],N=[3001], and Q=[2113]. Which of these matrices, if any, are similar to one another? Justify your answer without solving for a change-of-basis matrix P.
Compare the traces first; for the two matrices whose traces match, check whether they flatten space or are invertible.
Check the trace (sum of the main diagonal) of each matrix:
trace(M)trace(N)trace(Q)=2+2=4,=3+1=4,=2+3=5
Because trace(Q)=5=4,Q cannot be similar to M or N.
(Watch out: don't conclude that M and N are similar just because both have trace 4.) In M, the second column [42] is twice the first column [21], so M flattens the plane onto a line (rank 1, not invertible), whereas N is invertible (rank 2). Thus M and N cannot represent the same map either.
None of them are similar.
6
The matrices D=[3003] and T=[3013] both have trace 6, are both invertible with rank 2, and have the same determinant and eigenvalues. Are D and T similar? Prove your answer using the three-step detour P−1DP.
Write D=3I and simplify P−1(3I)P for any invertible matrix P.
(Watch out: passing all the quick number checks, with trace 6, rank 2, determinant 9, and eigenvalue 3, does not guarantee that D and T are similar.)
Because D=3I triples every vector in space, applying the three-step detour with any basis P gives:
P−1DP=P−1(3I)P=3P−1IP=3I=[3003]
Since D has the matrix [3003] in every basis, it can never equal T=[3013].