When a linear transformation moves vectors from one space to another, we often want to know whether we can undo it to get every starting vector back.
1
Undoing a Linear Transformation
For a linear transformation T:V→W that sends a vector v to an output w, its inverse transformationT−1:W→V is the rule that sends every output w right back to the single input v it came from.
What Makes a Linear Map Invertible
To reverse every arrow cleanly, T has to pass two tests at the same time:
only 0 maps to 0: Kernel={0}One-to-one (no squishing)andthe image fills all of WOnto (full coverage)
When you invert a matrix in Rn by setting up [A∣I] and row-reducing to [I∣A−1], the columns on the left are where the basis vectors landed, and row operations unscramble that stretched grid step-by-step back to the standard grid:
If row reduction hits a row of all zeros on the left, the matrix squished a dimension flat, so no inverse exists.
As long as the transformed basis vectors point in independent directions, T−1 unscrambles the stretched grid in W right back to the standard square grid in V. The exact instant one basis vector lands on the line of the other, the grid in W collapses flat onto a single line and fails both one-to-one and onto at once.
A transformation can be undone as long as the stretched grid covers the whole plane, and stops being reversible the instant the grid flattens onto a line.
2
Isomorphism: Same Space, Different Labels
Once we know when a linear map can be undone, we can compare two different-looking vector spaces. An isomorphism is simply an invertible linear transformation between two spaces V and W, telling us that both spaces work the exact same way under addition and scaling and only differ in the labels on their basis vectors.
Isomorphism: Same Space, Different Labels
In the space of polynomials up to degree 2, the powers 1,x, and x2 act as labels keeping three numbers from mixing during addition, doing the exact same job as the three vertical slots in R3:
aslot 1(1)+bslot 2(x)+cslot 3(x2)⟷abc
squish
not one-to-one
T
dimV=3
dimW=2
empty space
not onto
T
dimV=2
dimW=3
Because more starting directions than target directions forces a squish, and fewer starting directions leaves empty space, two spaces are isomorphic only when they have the exact same number of basis vectors:
dimV=dimW
3
Building an Isomorphism Between Spaces
For spaces like polynomials a+bx+cx2 or 2×2 matrices, the basis is easy to spot right away. What if a space hides its true dimension behind an equation, like the plane x+2y−z=0 inside R3, and we want to build an explicit isomorphism T:V→W between that plane and the space of polynomials of degree up to 1,a+bt?
Building an Isomorphism Between Spaces
x
y
z
V:x+2y−z=0
T
W:a+bt
isomorphic?
Text solution
Solve the equation to find the basis of VRewrite the plane equation x+2y−z=0 as z=x+2y and split a general vector on the plane by the free weights x and y to find its independent basis vectors.
Match dimensions and pair the two bases one-to-oneSince the plane V is built from 2 basis vectors and W is built from the 2 basis polynomials {1,t}, both spaces have dimension 2, and we build T by matching their basis vectors directly.
101⟷1,012⟷t
Keep the same basis weights going forward and backwardLinearity carries the weights x and y straight onto 1 and t to form T, while T−1 swaps the plane's basis vectors back in for 1 and t.
Txyz⟹T3513=x(1)+y(t)=x+yt=3+5t
T−1(3+5t)=3101+5012=3513
Every vector on the tilted plane x+2y−z=0 in R3 splits along the two basis directions 101 and 012 with weights x and y, matching the polynomial x+yt built from 1 and t with the exact same weights.
Every vector on this tilted 3D plane is built from just two basis directions, matching it one-to-one with the polynomial x + yt, built from the exact same two weights.
4
The Equal-Dimension Shortcut
Checking both one-to-one and onto separately takes two steps, so we want a faster way to test whether a linear map T:V→W between two spaces of the same dimension is invertible.
Remember from the Rank-Nullity Theorem that the dimension of the starting space acts as a fixed budget split between dimensions that get squished to zero and dimensions that survive into the image:
squished to 0dim(Kernel)+survives in Wdim(Image)=dim(Starting Space)
The Equal-Dimension Shortcut
V
W
T
one-to-one?
no squishing
onto?
full coverage
dimV=dimW
When the starting space and target space have the same dimension (dimV=dimW), checking that nothing gets squished (dim(Kernel)=0) forces the image to have the full dimension of W:
0+dim(Image)⟹dim(Image)=dimW=dimW
An image of dimension dimW inside W automatically fills the entire target space, giving onto for free.
5
Inverting a Map by Crossing to Rn
Now suppose we are given the calculus transformation T(p)=p(x)+p′(x) on polynomials up to degree 2 (a+bx+cx2), and we want to check whether T is invertible and find the starting polynomial p(x) that produces 5+3x+x2. Instead of guessing how to undo a derivative in polynomial space, we can use our coordinate isomorphism to cross into R3, invert the matrix there, and cross back.
Inverting a Map by Crossing to ℝⁿ
T(p)=p(x)+p′(x)
p(x)=?T5+3x+x2
invertible?
p(x)=4+x+x2
5+3x+x2
T
4
1
1
1
x
x2
5
3
1
1
x
x2
A−1
R3
labels off
labels on
Text solution
Plug the basis into T to build the matrix in R3Plug each vector from the basis {1,x,x2} into T(p)=p(x)+p′(x) and record the output coordinates as the columns of a 3×3 matrix.
Check that A is invertible and row-reduce next to the identityThe 1s down the diagonal, with 0s below them, show that no column is squished into the others (dim(Kernel)=0, so the equal-dimension shortcut guarantees T is invertible), and two row operations unscramble A into the identity to give A−1.
Multiply by A−1 and put the polynomial labels back onTranslate the target polynomial 5+3x+x2 into 531, multiply by A−1, and put the 1,x,x2 labels back onto the three slots.
Read the general calculus inverse from the columns of A−1Putting the {1,x,x2} labels back onto each column of A−1 shows that 1↦1,x↦−1+x, and x2↦2−2x+x2, which subtracts the first derivative and adds the second derivative.
T−1(q)=q(x)−q′(x)+q′′(x)
Key concepts
Everything here comes down to matching dimensions and tracking basis vectors. Before testing whether a map T:V→W is invertible, compare dimV and dimW by counting independent basis vectors, remembering to substitute any subspace equations first (so the plane z=x+2y in R3 has basis ⎩⎨⎧101,012⎭⎬⎫ and dimension 2, not 3). If dimV=dimW, stop immediately: you don't need to check the kernel or row-reduce, because unequal dimensions force either a squish or empty space, so no isomorphism can exist (for instance, polynomials of degree up to 2,a+bx+cx2, have dimension 3 and can never be isomorphic to 2×2 matrices of dimension 4).
When dimV=dimW, the spaces are isomorphic, and you can build an explicit isomorphism by pairing their basis vectors one-to-one (101↔1 and 012↔t, giving Txyz=x+yt). If you are handed a specific linear map T between two spaces of equal dimension, don't do the extra work of checking both one-to-one and onto: the Rank-Nullity budget dim(Kernel)+dim(Image)=dimV ties them together, so showing that only 0 maps to 0 (dim(Kernel)=0) guarantees full coverage automatically. Finally, to invert a map like T(p)=p(x)+p′(x), take off the basis labels {1,x,x2} to write the basis outputs as a matrix A in Rn, row-reduce [A∣I]→[I∣A−1] to unscramble the grid, and put the basis labels back on at the end.
6
Practice
1
Try to invert the matrices A=[1237] and B=[1236] by row-reducing each next to the identity matrix. Find A−1, and explain what the final left side of B's reduction tells you about what B does to the 2D grid.
Watch the sign on the second step: subtracting 3 times −2 gives 1−3(−2)=7; dropping the minus sign is a common slip that produces the wrong top-left entry −5 and [−5−2−31].
Start both reductions with the row operation that clears the 2 in the bottom-left corner, and watch whether the second column survives.
For A, set up [A∣I] and clear the entries off the diagonal:
(Watch the sign on the second step: subtracting 3 times −2 gives 1−3(−2)=7; dropping the minus sign is a common slip that produces the wrong top-left entry −5 and [−5−2−31].)
Reading the right side gives A−1=[7−2−31].
For B, applying the first row operation gives:
R2←R2−2R1[12361001][10301−201]
The row of all zeros on the left shows that B is not invertible: both basis columns [12] and [36] landed on the same line, squishing the 2D grid flat into a 1D line that row operations cannot un-squish.
2
Consider four vector spaces: U (polynomials of degree up to 2,a+bx+cx2), V (polynomials of degree up to 3,a+bx+cx2+dx3), W (2×2 matrices), and R3. Which pairs of spaces are isomorphic to each other, and why can there never be an isomorphism between U and W?
Count the number of independent basis vectors (or free number slots) in each space rather than looking at the highest exponent.
Count the basis vectors in each space:
U has the 3-vector basis {1,x,x2}, so dimU=3.
V has the 4-vector basis {1,x,x2,x3}, so dimV=4.
(Don't read the highest exponent as the dimension: forgetting the constant term 1 gives the wrong dimensions dimU=2 and dimV=3, which mistakenly pairs V with R3.)
W has 4 independent matrix slots, so dimW=4.
R3 has 3 coordinate slots, so dimR3=3.
Matching equal dimensions shows that U and R3 are isomorphic (both dimension 3, isomorphic to R3), and V and W are isomorphic (both dimension 4, isomorphic to R4). There can never be an isomorphism between U (dimension 3) and W (dimension 4) because any linear map from U to W can sweep out at most a 3-dimensional subspace inside 4-dimensional W (leaving empty space, so not onto), while any linear map from W to U must squish at least one of W's 4 independent directions flat to fit into 3-dimensional U (failing one-to-one).
3
Let T be a linear transformation from the space of 2×2 matrices to the space of polynomials of degree up to 3 (a+bx+cx2+dx3). Suppose you check the kernel of T and find that only the zero matrix [0000] maps to the zero polynomial. Without calculating the image of T, determine whether every polynomial of degree up to 3 gets hit by T and whether T is an isomorphism. Explain why using the Rank-Nullity Theorem.
Compare the dimensions of the two spaces first, then plug dim(Kernel) into the Rank-Nullity dimension budget.
Both the starting space (2×2 matrices with 4 slots) and the target space (polynomials up to degree 3 with basis {1,x,x2,x3}) have dimension 4. Since only the zero matrix maps to zero, no directions get squished, so T is one-to-one and:
dim(Kernel)=0
(Be careful not to write dim(Kernel)=1 just because the kernel contains one matrix: the zero matrix alone spans 0 directions, and using 1 gives the wrong image dimension 4−1=3 and the false conclusion that T is not onto.)
Plugging dim(Kernel)=0 into the Rank-Nullity Theorem gives:
Because a 4-dimensional image inside a 4-dimensional target space fills the entire space, every polynomial of degree up to 3 gets hit by T (onto) and T is an isomorphism.
4
Let V be the subspace of R3 made up of all vectors xyz satisfying 2x−y−z=0, and let W be the space of polynomials of degree up to 1,a+bt. Find a basis for V to show that V and W are isomorphic, build an explicit isomorphism T:V→W by pairing your basis with {1,t}, and compute both T27−3 and T−1(4−3t).
Solve the plane equation for z and substitute it into xyz to split the vector into x times one basis vector plus y times another.
Rewrite 2x−y−z=0 as z=2x−y and split a general vector on the plane by x and y:
xyz=xy2x−y=x102+y01−1
So ⎩⎨⎧102,01−1⎭⎬⎫ is a basis for V, giving dimV=2=dimW (with basis {1,t}), which shows V and W are isomorphic.
Pairing 102↔1 and 01−1↔t gives the explicit isomorphism and its inverse:
(Watch the double negative in the third slot: 2(4)−(−3)=11; adding instead of subtracting gives 8−3=5 and the wrong vector 4−35, which doesn't lie on the plane 2x−y−z=0.)
5
On the space of polynomials up to degree 2 (a+bx+cx2), define the linear transformation T(p)=p(x)−p′(x). Using the coordinate isomorphism with basis {1,x,x2}, write the 3×3 matrix for T in R3 and explain why T is invertible. Then invert the matrix to find the polynomial p(x) that produces the output 7−x+2x2, and read the columns of your inverse matrix to write a general calculus formula for T−1(q).
Plug 1,x, and x2 into p(x)−p′(x) one at a time, keeping the constant term in the top row and the x2 term in the bottom row.
Plug each basis polynomial from {1,x,x2} into T(p)=p(x)−p′(x) to form the columns of A:
The 1s down the diagonal, with 0s below them, show that no column is squished into the others (dim(Kernel)=0, one-to-one), and because both spaces have dimension 3, the equal-dimension shortcut guarantees T is onto and therefore invertible.
(Because A has negative entries above the diagonal, you must add rows to clear them, giving all positive entries in A−1=100110221; subtracting rows out of habit gives 100−1102−21 and the wrong polynomial 12−5x+2x2.)
Multiply A−1 by the coordinate vector 7−12 for 7−x+2x2: