Most matrices stretch, squish, or slant the grid when they transform space, turning squares into skewed parallelograms. When we want to rotate or flip space without changing a single length or angle, we need an orthogonal matrix.
1
Rigid Motions and Orthonormal Columns
An orthogonal matrix is a square matrix that moves the grid without stretching or slanting it. To see what its numbers must look like, we only need to track where the unit vectors (1,0) and (0,1) land, since those two arrows become the columns of the matrix.
the standard grid
a general matrix
an orthogonal matrix Q
Rigid Motions and Orthonormal Columns
For a square matrix Q=[c1c2⋯cn] to move the grid without stretching or slanting it, its columns must form an orthonormal set: every column has length 1, and different columns meet at 90∘:
length 1ci⋅ci=1and90∘ apartci⋅cj=0 for i=j
Even though "orthogonal" usually means 90∘ apart and "orthonormal" adds length 1, a square matrix with orthonormal columns is always called an orthogonal matrix.
2
Rotations, Reflections, and the Determinant
Once we pick the first column of a 2×2 orthogonal matrix on the unit circle, we want to find the possible second columns and tell from the numbers alone whether the matrix rotates the plane or flips it over.
Let's start with first column [5354] and see how the determinant separates the two choices.
Rotations, Reflections, and the Determinant
Text solution
Check that the first column has length 1Square and add the entries of c1 to check that it sits on the unit circle.
(53)2+(54)2=259+2516=1
Turn 90∘ both ways for the second columnSwap the two entries and make one of them negative to get the two unit vectors perpendicular to c1.
Qrot=[5354−5453],Qref=[535454−53]
Take the determinant to check for a flipBecause a 1×1 square keeps area 1, the determinant is +1 when the two axes rotate together and −1 when the grid flips over.
So Qrot is a pure counter-clockwise rotation, while Qref reflects the grid across a line.
As the second column moves around the unit circle while the first column stays fixed at (53,54), the transformed grid forms 1×1 squares at only the two spots 90∘ away from the first column. Moving from one spot to the other flips the square over and switches detQ from +1 to −1.
Even when both columns have length 1, the grid forms unskewed squares at only the two 90° positions: one a rotation, the other a reflection.
3
Why the Transpose Is the Inverse
Finding the inverse of a square matrix normally takes row reduction, but for an orthogonal matrix Q, its inverse is simply its transpose QT.
Transposing Q turns its columns c1,c2,c3 into rows, so multiplying QTQ takes every column-with-column dot product at once. Now let's see what happens when Q is square.
Why the Transpose Is the Inverse
Because each column dotted with itself is 1 and each column dotted with a different column is 0,QTQ becomes the identity matrix:
When a matrix is square, any matrix that multiplies it to give I works in both directions as its inverse:
transpose is the inverseQ−1=QTsoQTQ=IandQQT=I
While QTQ=I dots the columns of Q together, multiplying in the reverse order QQT=I dots the rows of Q together. This tells us that the rows of a square orthogonal matrix are automatically orthonormal too. (A tall 3×2 matrix with orthonormal columns still gives QTQ=I, but it has no inverse and QQT=I.)
When the two perpendicular unit columns of Q rotate by any angle, the rows of Q stay length 1 and 90∘ apart while rotating by the exact same angle in the opposite direction to undo Q.
As you rotate the orthonormal columns, the rows stay length 1 and at a right angle while turning the same angle the opposite way.
4
Solving Qx=b Without Elimination
Now let's solve the linear system Qx=b for the unknown vector x when Q is an orthogonal matrix.
Because Q−1=QT, we can skip elimination completely and find each coordinate of x with a single dot product.
Solving Qx = b Without Elimination
Here are the steps for solving [5354−5453][x1x2]=[55].
Text solution
Multiply both sides by QTReplace Q−1 with QT so the columns of Q become the rows multiplying b.
x=QTb=[53−545453][55]
Dot the first column of Q with bProject b onto c1 to get the first coordinate x1.
x1=c1⋅b=(53)(5)+(54)(5)=3+4=7
Dot the second column of Q with bProject b onto c2 to get the second coordinate x2.
x2=c2⋅b=(−54)(5)+(53)(5)=−4+3=−1
So x=[7−1], which means going 7 units along the first tilted axis and −1 unit along the second lands at (5,5).
5
Preserving Lengths, Dot Products, and Angles
When an orthogonal matrix Q transforms two vectors x and y into Qx and Qy, every coordinate changes, so let's see why their lengths, dot product, and angle stay the same.
Remember that the dot product of two column vectors can be written as a row-times-column product x⋅y=xTy, and that transposing a product reverses its order: (Qx)T=xTQT.
Preserving Lengths, Dot Products, and Angles
Writing the squared length ∥Qx∥2 as (Qx)⋅(Qx) puts QTQ right in the middle of the product, where it becomes I:
∥Qx∥2=(Qx)T(Qx)=xTIQTQx=xTIx=xTx=∥x∥2
Taking the square root gives ∥Qx∥=∥x∥. The exact same cancellation in the middle happens when we dot two transformed vectors Qx and Qy:
(Qx)⋅(Qy)=(Qx)T(Qy)=xTIQTQy=xTIy=xTy=x⋅y
Because the angle between two vectors depends only on their dot product and their lengths, the angle between Qx and Qy is always the same as the angle between x and y.
Across the entire plane, rotating or flipping any pair of vectors x and y into Qx and Qy changes their coordinates while keeping ∥Qx∥=∥x∥,(Qx)⋅(Qy)=x⋅y, and the angle between them unchanged.
Multiplying by an orthogonal matrix changes both vectors' coordinates while keeping their lengths, dot product, and the angle between them the same everywhere on the plane.
6
Multiplying Orthogonal Matrices
If we follow one orthogonal matrix Q2 with another orthogonal matrix Q1, is the combined matrix Q1Q2 also orthogonal?
Multiplying Orthogonal Matrices and Summary
To check with algebra that Q1Q2 has orthonormal columns, we compute (Q1Q2)T(Q1Q2) by reversing the order inside the transpose so Q1T and Q1 meet in the middle:
Because (Q1Q2)T(Q1Q2)=I, the product of two orthogonal matrices is always orthogonal, with inverse (Q1Q2)−1=(Q1Q2)T=Q2TQ1T.
Key concepts
Everything with an orthogonal matrix Q comes back to one fact: Q is a square matrix whose columns (and rows) have length 1 and are 90∘ apart, so QTQ=I and QQT=I.
Testing or completing Q
First check that the matrix is square: a tall matrix like D=100010 has orthonormal columns (DTD=I), but it is not an orthogonal matrix and cannot be inverted with DT. Next check both parts of "orthonormal": perpendicular columns like [11−11] have dot product 0, but their lengths are 2 instead of 1, so the matrix is not orthogonal unless each column is divided by 2. When finding unknown entries in Q, use length 1 (ri⋅ri=1) on a row or column with a single positive unknown, and use dot product 0 (ri⋅rj=0) whenever you need the unknown's sign.
Rotation vs. reflection
Once you know Q is orthogonal, check detQ. A determinant of +1 means a pure rotation (like [5354−5453]), while −1 means a reflection is included (like [535454−53]).
Inverting Q and solving Qx=b
Never row reduce an orthogonal matrix. Swap rows and columns to get Q−1=QT, and solve Qx=b by dotting each column of Q with b:xi=ci⋅b. (If the columns are not both length 1 and perpendicular, QT is not the inverse, so QTb does not solve the system.)
Lengths, dot products, and products of Q
When an orthogonal matrix appears inside ∥Qx∥ or (Qx)⋅(Qy), don't multiply Qx: drop Q and compute ∥x∥ or x⋅y directly. For a product of orthogonal matrices Q1Q2, reverse the order when transposing to get (Q1Q2)−1=Q2TQ1T.
7
Practice
1
Which of the matrices A=[100−1],B=[11−11],C=[105354], and D=100010 are orthogonal, and why do the others fail?
Check the shape of each matrix first, then test both parts of "orthonormal": do the columns dot to 0, and does each column have length 1?
For A, the matrix is 2×2, its columns have lengths 12+02=1 and 02+(−1)2=1, and their dot product is (1)(0)+(0)(−1)=0, so A is orthogonal.
For B, the columns dot to (1)(−1)+(1)(1)=0, which makes it tempting to call B orthogonal right away, so be careful not to forget the length check, which gives 12+12=2=1.
For C, both columns have length 1, but their dot product is (1)(53)+(0)(54)=53=0.
For D, the columns are orthonormal (DTD=I), which often tricks people into counting it, but a 3×2 matrix is not square (DDT=I).
Only A is orthogonal; B fails because its columns have length 2 instead of 1,C fails because its columns have dot product 53 instead of 0, and D fails because it is not square.
2
A 2×2 orthogonal matrix Q has first column [1351312]. Find both possible second columns, and state which matrix is a pure rotation and which is a reflection.
Making both entries negative without swapping gives [−135−1312], which points 180∘ opposite the first column and gives dot product −1 instead of 0.
To turn a 2D vector by 90∘ without changing its length, what do you do to its two entries? Then check the sign of the determinant for each matrix.
Swapping the entries of [1351312] and making one entry negative gives the two unit vectors perpendicular to the first column: [−1312135] and [1312−135]. (Making both entries negative without swapping gives [−135−1312], which points 180∘ opposite the first column and gives dot product −1 instead of 0.)
Now take the determinant ad−bc of each matrix (be careful not to drop the minus sign in −bc, which would give 16925−169144=−169119 instead of ±1):
The two possible second columns are [−1312135] (giving Q=[1351312−1312135], a pure rotation with detQ=+1) and [1312−135] (giving Q=[13513121312−135], a reflection with detQ=−1).
3
Given the orthogonal matrix Q=[5453−5354] and b=[10−5], write down Q−1 and solve Qx=b using dot products without row reduction.
Once you swap rows and columns to get Q−1=QT, are you dotting b with the rows of Q or the columns of Q?
Transposing Q gives the inverse:
Q−1=QT=[54−535354]
To compute x=QTb, dot the columns of Q (the rows of QT) with b. Don't dot the original rows of Q with b, which computes Qb and gives the wrong vector [112]:
x1=c1⋅b=(54)(10)+(53)(−5)=8−3=5
x2=c2⋅b=(−53)(10)+(54)(−5)=−6−4=−10
Q−1=[54−535354],x=[5−10]
4
Find the missing entries a,b, and c in the 3×3 orthogonal matrix Q=3132−3232b32a−32c given that a>0.
Use length 1 on the row that has only a missing. Once you have a, use dot products equal to 0 rather than lengths so you don't lose the signs of b and c.
Row 1 has length 1 and a>0:
(31)2+(32)2+a2⟹95+a2⟹a=1=1=32
To find b and c, don't use the lengths of Column 2 or Column 3: squaring the unknown gives b2=91 and c2=91, which hides the minus sign and leads to the wrong sign c=31. Instead, dot perpendicular columns or rows to 0:
Let Q be an unknown 3×3 orthogonal matrix, and let x=12−2 and y=212. Without knowing the entries of Q, find ∥Qx∥,(Qx)⋅(Qy), and the angle between Qx and Qy.
Remember to take the square root at the end rather than stopping at ∥x∥2=9.
What happens to QTQ inside a length or a dot product?
Because Q keeps lengths the same, ∥Qx∥=∥x∥ (remember to take the square root at the end rather than stopping at ∥x∥2=9):
∥Qx∥=∥x∥=12+22+(−2)2=9=3
Because Q keeps dot products the same, (Qx)⋅(Qy)=x⋅y:
(Qx)⋅(Qy)=(1)(2)+(2)(1)+(−2)(2)=2+2−4=0
A dot product of 0 means cosθ=0, so the vectors are perpendicular (don't confuse a dot product of 0 with an angle of 0∘).
∥Qx∥=3,(Qx)⋅(Qy)=0,angle=90∘
6
Suppose Q1 and Q2 are n×n orthogonal matrices where Q1 is a rotation (detQ1=+1) and Q2 is a reflection (detQ2=−1). Show algebraically that M=Q1Q2T is orthogonal, find M−1 in terms of Q1 and Q2, and determine whether detM is +1 or −1.
When you transpose the product Q1Q2T to compute MTM, what happens to the order of the two matrices?
Reverse the order of the matrices when transposing M=Q1Q2T. Transposing in place without reversing the order gives the wrong expression Q1TQ2, where the middle matrices cannot cancel:
MT=(Q1Q2T)T=(Q2T)TQ1T=Q2Q1T
Multiply MTM and use Q1TQ1=I followed by Q2Q2T=I:
This shows M is orthogonal, so its inverse is M−1=MT=Q2Q1T. Finally, transposing Q2 does not change its determinant (detQ2T=−1, not +1), so following the reflection Q2T with the rotation Q1 leaves space flipped: