Linear AlgebraOrthogonality and Projections9 of 9
Orthogonality and Projections · 9

Orthogonal Matrices

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)\first{(1, 0)} and (0,1)\second{(0, 1)} land, since those two arrows become the columns of the matrix.

(1,0)\htmlClass{lc-green}{(1, 0)}
(0,1)\htmlClass{lc-pink}{(0, 1)}
the standard grid
c1\htmlClass{lc-green}{c_1}
c2\htmlClass{lc-pink}{c_2}
a general matrix
c1\htmlClass{lc-green}{c_1}
c2\htmlClass{lc-pink}{c_2}
11
11
an orthogonal matrix QQ
Rigid Motions and Orthonormal Columns

For a square matrix Q=[c1c2⋯cn]Q = \begin{bmatrix} \first{c_1} & \second{c_2} & \cdots & c_n \end{bmatrix} to move the grid without stretching or slanting it, its columns must form an orthonormal set: every column has length 11, and different columns meet at 90∘90^\circ:

ci⋅ci=1⏟length 1andci⋅cj=0 for i≠j⏟90∘ apart\displaystyle \begin{aligned} & \underbrace{\first{c_i} \cdot \first{c_i} = 1}_{\text{length } 1} \qquad \text{and} \\ &\underbrace{\first{c_i} \cdot \second{c_j} = 0 \text{ for } i \neq j}_{90^\circ \text{ apart}} \end{aligned}

Even though "orthogonal" usually means 90∘90^\circ apart and "orthonormal" adds length 11, 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 2×22 \times 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 [3545]\first{\begin{bmatrix} \tfrac{3}{5} \\ \tfrac{4}{5} \end{bmatrix}} and see how the determinant separates the two choices.

Rotations, Reflections, and the Determinant
[3545]\htmlClass{lc-green}{\begin{bmatrix} \tfrac{3}{5} \\[6pt] \tfrac{4}{5} \end{bmatrix}}
1\htmlClass{lc-mute}{1}
1\htmlClass{lc-mute}{1}
Q=[35?45?]Q = \begin{bmatrix} \htmlClass{lc-green}{\tfrac{3}{5}} & \htmlClass{lc-pink}{?} \\[6pt] \htmlClass{lc-green}{\tfrac{4}{5}} & \htmlClass{lc-pink}{?} \end{bmatrix}
det⁡Q=  ?\det Q = \;?
Text solution
Check that the first column has length 11Square and add the entries of c1\first{c_1} to check that it sits on the unit circle.
(35)2+(45)2\displaystyle \left(\first{\frac{3}{5}}\right)^2 + \left(\first{\frac{4}{5}}\right)^2 =925+1625\displaystyle = \frac{9}{25} + \frac{16}{25} =1\displaystyle = 1
Turn 90∘90^\circ both ways for the second columnSwap the two entries and make one of them negative to get the two unit vectors perpendicular to c1\first{c_1}.
Qrot=[35−454535],Qref=[354545−35]\displaystyle Q_{\text{rot}} = \begin{bmatrix} \first{\tfrac{3}{5}} & \second{-\tfrac{4}{5}} \\ \first{\tfrac{4}{5}} & \second{\tfrac{3}{5}} \end{bmatrix}, \qquad Q_{\text{ref}} = \begin{bmatrix} \first{\tfrac{3}{5}} & \second{\tfrac{4}{5}} \\ \first{\tfrac{4}{5}} & \second{-\tfrac{3}{5}} \end{bmatrix}
Take the determinant to check for a flipBecause a 1×11 \times 1 square keeps area 11, the determinant is +1+1 when the two axes rotate together and −1-1 when the grid flips over.
det⁡Qrot=(35)(35)−(−45)(45)=925+1625=+1det⁡Qref=(35)(−35)−(45)(45)=−925−1625=−1\displaystyle \begin{aligned} \det Q_{\text{rot}} &= \left(\first{\frac{3}{5}}\right)\left(\second{\frac{3}{5}}\right) \\ &\quad - \left(\second{-\frac{4}{5}}\right)\left(\first{\frac{4}{5}}\right) \\ &= \frac{9}{25} + \frac{16}{25} = +1 \\[10pt] \det Q_{\text{ref}} &= \left(\first{\frac{3}{5}}\right)\left(\second{-\frac{3}{5}}\right) \\ &\quad - \left(\second{\frac{4}{5}}\right)\left(\first{\frac{4}{5}}\right) \\ &= -\frac{9}{25} - \frac{16}{25} = -1 \end{aligned}

So QrotQ_{\text{rot}} is a pure counter-clockwise rotation, while QrefQ_{\text{ref}} reflects the grid across a line.

As the second column moves around the unit circle while the first column stays fixed at (35,45)\first{(\tfrac{3}{5}, \tfrac{4}{5})}, the transformed grid forms 1×11 \times 1 squares at only the two spots 90∘90^\circ away from the first column. Moving from one spot to the other flips the square over and switches det⁡Q\det Q from +1+1 to −1-1.

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 QQ, its inverse is simply its transpose QTQ^T.

Transposing QQ turns its columns c1,c2,c3\first{c_1}, \second{c_2}, \third{c_3} into rows, so multiplying QTQQ^T Q takes every column-with-column dot product at once. Now let's see what happens when QQ is square.

Why the Transpose Is the Inverse
c1\htmlClass{lc-green}{c_1}
c2\htmlClass{lc-pink}{c_2}
Q=[35−454535]Q = \begin{bmatrix} \htmlClass{lc-green}{\tfrac{3}{5}} & \htmlClass{lc-pink}{-\tfrac{4}{5}} \\[6pt] \htmlClass{lc-green}{\tfrac{4}{5}} & \htmlClass{lc-pink}{\tfrac{3}{5}} \end{bmatrix}
Q−1=  ?Q^{-1} = \;?

Because each column dotted with itself is 11 and each column dotted with a different column is 00, QTQQ^T Q becomes the identity matrix:

QTQ=[c1⋅c1c1⋅c2c1⋅c3c2⋅c1c2⋅c2c2⋅c3c3⋅c1c3⋅c2c3⋅c3]=[100010001]=I\displaystyle \begin{aligned} Q^T Q &= \begin{bmatrix} \first{c_1} \cdot \first{c_1} & \first{c_1} \cdot \second{c_2} & \first{c_1} \cdot \third{c_3} \\ \second{c_2} \cdot \first{c_1} & \second{c_2} \cdot \second{c_2} & \second{c_2} \cdot \third{c_3} \\ \third{c_3} \cdot \first{c_1} & \third{c_3} \cdot \second{c_2} & \third{c_3} \cdot \third{c_3} \end{bmatrix} \\ &= \begin{bmatrix} 1 & 0 & 0 \\ 0 & 1 & 0 \\ 0 & 0 & 1 \end{bmatrix} = I \end{aligned}

When a matrix is square, any matrix that multiplies it to give II works in both directions as its inverse:

Q−1=QT⏟transpose is the inversesoQTQ\displaystyle \underbrace{Q^{-1} = Q^T}_{\text{transpose is the inverse}} \qquad \text{so} \qquad Q^T Q =IandQQT\displaystyle = I \quad \text{and} \quad Q Q^T =I\displaystyle = I

While QTQ=IQ^T Q = I dots the columns of QQ together, multiplying in the reverse order QQT=IQ Q^T = I dots the rows of QQ together. This tells us that the rows of a square orthogonal matrix are automatically orthonormal too. (A tall 3×23 \times 2 matrix with orthonormal columns still gives QTQ=IQ^T Q = I, but it has no inverse and QQT≠IQ Q^T \fails{\neq I}.)

When the two perpendicular unit columns of QQ rotate by any angle, the rows of QQ stay length 11 and 90∘90^\circ apart while rotating by the exact same angle in the opposite direction to undo QQ.

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=bQx = b Without Elimination

Now let's solve the linear system Qx=bQ\target{x} = \target{b} for the unknown vector x\target{x} when QQ is an orthogonal matrix.

Because Q−1=QTQ^{-1} = Q^T, we can skip elimination completely and find each coordinate of x\target{x} with a single dot product.

c1\htmlClass{lc-green}{c_1}
c2\htmlClass{lc-pink}{c_2}
b=(5,5)\htmlClass{lc-blue}{b} = \htmlClass{lc-blue}{(5, 5)}
x1=7\htmlClass{lc-green}{x_1} = \htmlClass{lc-green}{7}
x2=−1\htmlClass{lc-pink}{x_2} = \htmlClass{lc-pink}{-1}
Solving Qx = b Without Elimination
b\htmlClass{lc-blue}{b}
5\htmlClass{lc-mute}{5}
5\htmlClass{lc-mute}{5}
[35−454535]Qx=[55]b\overset{Q}{\begin{bmatrix} \htmlClass{lc-green}{\tfrac{3}{5}} & \htmlClass{lc-pink}{-\tfrac{4}{5}} \\[6pt] \htmlClass{lc-green}{\tfrac{4}{5}} & \htmlClass{lc-pink}{\tfrac{3}{5}} \end{bmatrix}}\htmlClass{lc-blue}{x} = \overset{\htmlClass{lc-blue}{b}}{\htmlClass{lc-blue}{\begin{bmatrix} 5 \\ 5 \end{bmatrix}}}
x=  ?\htmlClass{lc-blue}{x} = \;?

Here are the steps for solving [35−454535][x1x2]=[55]\begin{bmatrix} \first{\tfrac{3}{5}} & \second{-\tfrac{4}{5}} \\ \first{\tfrac{4}{5}} & \second{\tfrac{3}{5}} \end{bmatrix} \begin{bmatrix} \first{x_1} \\ \second{x_2} \end{bmatrix} = \target{\begin{bmatrix} 5 \\ 5 \end{bmatrix}}.

Text solution
Multiply both sides by QTQ^TReplace Q−1Q^{-1} with QTQ^T so the columns of QQ become the rows multiplying b\target{b}.
x=QTb=[3545−4535][55]\displaystyle \target{x} = Q^T \target{b} = \begin{bmatrix} \first{\tfrac{3}{5}} & \first{\tfrac{4}{5}} \\ \second{-\tfrac{4}{5}} & \second{\tfrac{3}{5}} \end{bmatrix} \target{\begin{bmatrix} 5 \\ 5 \end{bmatrix}}
Dot the first column of QQ with b\target{b}Project b\target{b} onto c1\first{c_1} to get the first coordinate x1\first{x_1}.
x1\displaystyle \first{x_1} =c1⋅b\displaystyle = \first{c_1} \cdot \target{b} =(35)(5)+(45)(5)\displaystyle = \left(\first{\frac{3}{5}}\right)(\target{5}) + \left(\first{\frac{4}{5}}\right)(\target{5}) =3+4\displaystyle = 3 + 4 =7\displaystyle = \first{7}
Dot the second column of QQ with b\target{b}Project b\target{b} onto c2\second{c_2} to get the second coordinate x2\second{x_2}.
x2\displaystyle \second{x_2} =c2⋅b\displaystyle = \second{c_2} \cdot \target{b} =(−45)(5)+(35)(5)\displaystyle = \left(\second{-\frac{4}{5}}\right)(\target{5}) + \left(\second{\frac{3}{5}}\right)(\target{5}) =−4+3\displaystyle = -4 + 3 =−1\displaystyle = \second{-1}

So x=[7−1]\target{x} = \begin{bmatrix} \first{7} \\ \second{-1} \end{bmatrix}, which means going 7\first{7} units along the first tilted axis and −1\second{-1} unit along the second lands at (5,5)\target{(5, 5)}.

5

Preserving Lengths, Dot Products, and Angles

When an orthogonal matrix QQ transforms two vectors x\target{x} and yy into QxQ\target{x} and QyQy, 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\target{x} \cdot y = \target{x}^T y, and that transposing a product reverses its order: (Qx)T=xTQT(Q\target{x})^T = \target{x}^T Q^T.

Preserving Lengths, Dot Products, and Angles
x\htmlClass{lc-blue}{x}
QxQ\htmlClass{lc-blue}{x}
∥Qx∥=  ?\|Q\htmlClass{lc-blue}{x}\| = \;?

Writing the squared length ∥Qx∥2\|Q\target{x}\|^2 as (Qx)⋅(Qx)(Q\target{x}) \cdot (Q\target{x}) puts QTQQ^T Q right in the middle of the product, where it becomes II:

∥Qx∥2=(Qx)T(Qx)=xTQTQ⏟Ix=xTIx=xTx=∥x∥2\displaystyle \begin{aligned} \|Q\target{x}\|^2 &= (Q\target{x})^T(Q\target{x}) = \target{x}^T \underbrace{Q^T Q}_{I} \target{x} \\ &= \target{x}^T I \target{x} = \target{x}^T \target{x} = \|\target{x}\|^2 \end{aligned}

Taking the square root gives ∥Qx∥=∥x∥\|Q\target{x}\| = \|\target{x}\|. The exact same cancellation in the middle happens when we dot two transformed vectors QxQ\target{x} and QyQy:

(Qx)⋅(Qy)=(Qx)T(Qy)=xTQTQ⏟Iy=xTIy=xTy=x⋅y\displaystyle \begin{aligned} (Q\target{x}) \cdot (Qy) &= (Q\target{x})^T(Qy) = \target{x}^T \underbrace{Q^T Q}_{I} y \\ &= \target{x}^T I y = \target{x}^T y = \target{x} \cdot y \end{aligned}

Because the angle between two vectors depends only on their dot product and their lengths, the angle between QxQ\target{x} and QyQy is always the same as the angle between x\target{x} and yy.

Across the entire plane, rotating or flipping any pair of vectors x\target{x} and yy into QxQ\target{x} and QyQy changes their coordinates while keeping ∥Qx∥=∥x∥\|Q\target{x}\| = \|\target{x}\|, (Qx)⋅(Qy)=x⋅y(Q\target{x}) \cdot (Qy) = \target{x} \cdot 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 Q2Q_2 with another orthogonal matrix Q1Q_1, is the combined matrix Q1Q2Q_1 Q_2 also orthogonal?

Multiplying Orthogonal Matrices and Summary
11
11
Q1Q2⏟orthogonal?\underbrace{Q_1 Q_2}_{\text{orthogonal?}}

To check with algebra that Q1Q2Q_1 Q_2 has orthonormal columns, we compute (Q1Q2)T(Q1Q2)(Q_1 Q_2)^T (Q_1 Q_2) by reversing the order inside the transpose so Q1TQ_1^T and Q1Q_1 meet in the middle:

(Q1Q2)T(Q1Q2)=Q2TQ1TQ1⏟IQ2=Q2TIQ2=Q2TQ2=I\displaystyle \begin{aligned} (Q_1 Q_2)^T (Q_1 Q_2) &= Q_2^T \underbrace{Q_1^T Q_1}_{I} Q_2 \\ &= Q_2^T I Q_2 = Q_2^T Q_2 = I \end{aligned}

Because (Q1Q2)T(Q1Q2)=I(Q_1 Q_2)^T (Q_1 Q_2) = I, the product of two orthogonal matrices is always orthogonal, with inverse (Q1Q2)−1=(Q1Q2)T=Q2TQ1T(Q_1 Q_2)^{-1} = (Q_1 Q_2)^T = Q_2^T Q_1^T.

Key concepts

Everything with an orthogonal matrix QQ comes back to one fact: QQ is a square matrix whose columns (and rows) have length 11 and are 90∘90^\circ apart, so QTQ=IQ^T Q = I and QQT=IQ Q^T = I.

Testing or completing QQ

First check that the matrix is square: a tall matrix like D=[100100]D = \begin{bmatrix} \first{1} & \second{0} \\ \first{0} & \second{1} \\ \first{0} & \second{0} \end{bmatrix} has orthonormal columns (DTD=ID^T D = I), but it is not an orthogonal matrix and cannot be inverted with DTD^T. Next check both parts of "orthonormal": perpendicular columns like [1−111]\begin{bmatrix} \first{1} & \second{-1} \\ \first{1} & \second{1} \end{bmatrix} have dot product 00, but their lengths are 2\fails{\sqrt{2}} instead of 11, so the matrix is not orthogonal unless each column is divided by 2\sqrt{2}. When finding unknown entries in QQ, use length 11 (ri⋅ri=1r_i \cdot r_i = 1) on a row or column with a single positive unknown, and use dot product 00 (ri⋅rj=0r_i \cdot r_j = 0) whenever you need the unknown's sign.

Rotation vs. reflection

Once you know QQ is orthogonal, check det⁡Q\det Q. A determinant of +1+1 means a pure rotation (like [35−454535]\begin{bmatrix} \first{\tfrac{3}{5}} & \second{-\tfrac{4}{5}} \\ \first{\tfrac{4}{5}} & \second{\tfrac{3}{5}} \end{bmatrix}), while −1-1 means a reflection is included (like [354545−35]\begin{bmatrix} \first{\tfrac{3}{5}} & \second{\tfrac{4}{5}} \\ \first{\tfrac{4}{5}} & \second{-\tfrac{3}{5}} \end{bmatrix}).

Inverting QQ and solving Qx=bQ\target{x} = \target{b}

Never row reduce an orthogonal matrix. Swap rows and columns to get Q−1=QTQ^{-1} = Q^T, and solve Qx=bQ\target{x} = \target{b} by dotting each column of QQ with b\target{b}: xi=ci⋅b\first{x_i} = \first{c_i} \cdot \target{b}. (If the columns are not both length 11 and perpendicular, QTQ^T is not the inverse, so QTbQ^T \target{b} does not solve the system.)

Lengths, dot products, and products of QQ

When an orthogonal matrix appears inside ∥Qx∥\|Q\target{x}\| or (Qx)⋅(Qy)(Q\target{x}) \cdot (Qy), don't multiply QxQ\target{x}: drop QQ and compute ∥x∥\|\target{x}\| or x⋅y\target{x} \cdot y directly. For a product of orthogonal matrices Q1Q2Q_1 Q_2, reverse the order when transposing to get (Q1Q2)−1=Q2TQ1T(Q_1 Q_2)^{-1} = Q_2^T Q_1^T.

7

Practice

1

Which of the matrices A=[100−1]A = \begin{bmatrix} \first{1} & \second{0} \\ \first{0} & \second{-1} \end{bmatrix}, B=[1−111]B = \begin{bmatrix} \first{1} & \second{-1} \\ \first{1} & \second{1} \end{bmatrix}, C=[135045]C = \begin{bmatrix} \first{1} & \second{\tfrac{3}{5}} \\ \first{0} & \second{\tfrac{4}{5}} \end{bmatrix}, and D=[100100]D = \begin{bmatrix} \first{1} & \second{0} \\ \first{0} & \second{1} \\ \first{0} & \second{0} \end{bmatrix} 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 00, and does each column have length 11?

For AA, the matrix is 2×22 \times 2, its columns have lengths 12+02=1\sqrt{\first{1}^2 + \first{0}^2} = 1 and 02+(−1)2=1\sqrt{\second{0}^2 + (\second{-1})^2} = 1, and their dot product is (1)(0)+(0)(−1)=0(\first{1})(\second{0}) + (\first{0})(\second{-1}) = 0, so AA is orthogonal.
For BB, the columns dot to (1)(−1)+(1)(1)=0(\first{1})(\second{-1}) + (\first{1})(\second{1}) = 0, which makes it tempting to call BB orthogonal right away, so be careful not to forget the length check, which gives 12+12=2≠1\sqrt{\first{1}^2 + \first{1}^2} = \fails{\sqrt{2} \neq 1}.
For CC, both columns have length 11, but their dot product is (1)(35)+(0)(45)=35≠0(\first{1})(\second{\tfrac{3}{5}}) + (\first{0})(\second{\tfrac{4}{5}}) = \fails{\tfrac{3}{5} \neq 0}.
For DD, the columns are orthonormal (DTD=ID^T D = I), which often tricks people into counting it, but a 3×23 \times 2 matrix is not square (DDT≠ID D^T \fails{\neq I}).
Only AA is orthogonal; BB fails because its columns have length 2\fails{\sqrt{2}} instead of 11, CC fails because its columns have dot product 35\fails{\tfrac{3}{5}} instead of 00, and DD fails because it is not square.
2

A 2×22 \times 2 orthogonal matrix QQ has first column [5131213]\first{\begin{bmatrix} \tfrac{5}{13} \\ \tfrac{12}{13} \end{bmatrix}}. Find both possible second columns, and state which matrix is a pure rotation and which is a reflection.

To turn a 2D vector by 90∘90^\circ 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 [5131213]\first{\begin{bmatrix} \tfrac{5}{13} \\ \tfrac{12}{13} \end{bmatrix}} and making one entry negative gives the two unit vectors perpendicular to the first column: [−1213513]\second{\begin{bmatrix} -\tfrac{12}{13} \\ \tfrac{5}{13} \end{bmatrix}} and [1213−513]\second{\begin{bmatrix} \tfrac{12}{13} \\ -\tfrac{5}{13} \end{bmatrix}}. (Making both entries negative without swapping gives [−513−1213]\fails{\begin{bmatrix} -\tfrac{5}{13} \\ -\tfrac{12}{13} \end{bmatrix}}, which points 180∘180^\circ opposite the first column and gives dot product −1\fails{-1} instead of 00.)
Now take the determinant ad−bcad - bc of each matrix (be careful not to drop the minus sign in −bc-bc, which would give 25169−144169=−119169\tfrac{25}{169} - \tfrac{144}{169} = \fails{-\tfrac{119}{169}} instead of ±1\pm 1):
det⁡[513−12131213513]=25169−(−144169)=+1det⁡[51312131213−513]=−25169−144169=−1\displaystyle \begin{aligned} \det \begin{bmatrix} \first{\tfrac{5}{13}} & \second{-\tfrac{12}{13}} \\[4pt] \first{\tfrac{12}{13}} & \second{\tfrac{5}{13}} \end{bmatrix} &= \frac{25}{169} - \left(-\frac{144}{169}\right) \\ &= +1 \\[10pt] \det \begin{bmatrix} \first{\tfrac{5}{13}} & \second{\tfrac{12}{13}} \\[4pt] \first{\tfrac{12}{13}} & \second{-\tfrac{5}{13}} \end{bmatrix} &= -\frac{25}{169} - \frac{144}{169} \\ &= -1 \end{aligned}
The two possible second columns are [−1213513]\second{\begin{bmatrix} -\tfrac{12}{13} \\ \tfrac{5}{13} \end{bmatrix}} (giving Q=[513−12131213513]Q = \begin{bmatrix} \first{\tfrac{5}{13}} & \second{-\tfrac{12}{13}} \\ \first{\tfrac{12}{13}} & \second{\tfrac{5}{13}} \end{bmatrix}, a pure rotation with det⁡Q=+1\det Q = +1) and [1213−513]\second{\begin{bmatrix} \tfrac{12}{13} \\ -\tfrac{5}{13} \end{bmatrix}} (giving Q=[51312131213−513]Q = \begin{bmatrix} \first{\tfrac{5}{13}} & \second{\tfrac{12}{13}} \\ \first{\tfrac{12}{13}} & \second{-\tfrac{5}{13}} \end{bmatrix}, a reflection with det⁡Q=−1\det Q = -1).
3

Given the orthogonal matrix Q=[45−353545]Q = \begin{bmatrix} \first{\tfrac{4}{5}} & \second{-\tfrac{3}{5}} \\ \first{\tfrac{3}{5}} & \second{\tfrac{4}{5}} \end{bmatrix} and b=[10−5]\target{b} = \target{\begin{bmatrix} 10 \\ -5 \end{bmatrix}}, write down Q−1Q^{-1} and solve Qx=bQ\target{x} = \target{b} using dot products without row reduction.

Once you swap rows and columns to get Q−1=QTQ^{-1} = Q^T, are you dotting b\target{b} with the rows of QQ or the columns of QQ?

Transposing QQ gives the inverse:
Q−1=QT=[4535−3545]\displaystyle Q^{-1} = Q^T = \begin{bmatrix} \first{\tfrac{4}{5}} & \first{\tfrac{3}{5}} \\ \second{-\tfrac{3}{5}} & \second{\tfrac{4}{5}} \end{bmatrix}
To compute x=QTb\target{x} = Q^T \target{b}, dot the columns of QQ (the rows of QTQ^T) with b\target{b}. Don't dot the original rows of QQ with b\target{b}, which computes QbQ\target{b} and gives the wrong vector [112]\fails{\begin{bmatrix} 11 \\ 2 \end{bmatrix}}:
x1\displaystyle \first{x_1} =c1⋅b\displaystyle = \first{c_1} \cdot \target{b} =(45)(10)+(35)(−5)\displaystyle = \left(\first{\frac{4}{5}}\right)(\target{10}) + \left(\first{\frac{3}{5}}\right)(\target{-5}) =8−3\displaystyle = 8 - 3 =5\displaystyle = \first{5}
x2\displaystyle \second{x_2} =c2⋅b\displaystyle = \second{c_2} \cdot \target{b} =(−35)(10)+(45)(−5)\displaystyle = \left(\second{-\frac{3}{5}}\right)(\target{10}) + \left(\second{\frac{4}{5}}\right)(\target{-5}) =−6−4\displaystyle = -6 - 4 =−10\displaystyle = \second{-10}
Q−1=[4535−3545],x=[5−10]\displaystyle Q^{-1} = \begin{bmatrix} \first{\tfrac{4}{5}} & \first{\tfrac{3}{5}} \\ \second{-\tfrac{3}{5}} & \second{\tfrac{4}{5}} \end{bmatrix}, \qquad \target{x} = \begin{bmatrix} \first{5} \\ \second{-10} \end{bmatrix}
4

Find the missing entries a\third{a}, b\second{b}, and c\third{c} in the 3×33 \times 3 orthogonal matrix Q=[1323a23b−23−2323c]Q = \begin{bmatrix} \first{\tfrac{1}{3}} & \second{\tfrac{2}{3}} & \third{a} \\ \first{\tfrac{2}{3}} & \second{b} & \third{-\tfrac{2}{3}} \\ \first{-\tfrac{2}{3}} & \second{\tfrac{2}{3}} & \third{c} \end{bmatrix} given that a>0\third{a} > 0.

Use length 11 on the row that has only a\third{a} missing. Once you have a\third{a}, use dot products equal to 00 rather than lengths so you don't lose the signs of b\second{b} and c\third{c}.

Row 1 has length 11 and a>0\third{a} > 0:
(13)2+(23)2+a2=1  ⟹  59+a2=1  ⟹  a=23\displaystyle \begin{aligned} \left(\first{\frac{1}{3}}\right)^2 + \left(\second{\frac{2}{3}}\right)^2 + \third{a}^2 &= 1 \\ \implies \frac{5}{9} + \third{a}^2 &= 1 \\ \implies \third{a} &= \frac{2}{3} \end{aligned}
To find b\second{b} and c\third{c}, don't use the lengths of Column 2 or Column 3: squaring the unknown gives b2=19\second{b}^2 = \tfrac{1}{9} and c2=19\third{c}^2 = \tfrac{1}{9}, which hides the minus sign and leads to the wrong sign c=13\third{c} = \fails{\tfrac{1}{3}}. Instead, dot perpendicular columns or rows to 00:
Column 1⋅Column 2=(13)(23)+(23)b+(−23)(23)=0  ⟹  29+23b−49=0  ⟹  b=13\displaystyle \begin{aligned} & \first{\text{Column 1}} \cdot \second{\text{Column 2}} \\ &\quad = \left(\first{\frac{1}{3}}\right)\left(\second{\frac{2}{3}}\right) \\ &\quad + \left(\first{\frac{2}{3}}\right)\second{b}+ \left(\first{-\frac{2}{3}}\right)\left(\second{\frac{2}{3}}\right) = 0 \\ & \implies \frac{2}{9} + \frac{2}{3}\second{b} - \frac{4}{9} = 0 \\ & \implies \second{b} = \frac{1}{3} \end{aligned}
Row 1⋅Row 3=(13)(−23)+(23)(23)+(23)c=0  ⟹  −29+49+23c=0  ⟹  c=−13\displaystyle \begin{aligned} & \text{Row 1} \cdot \text{Row 3} = \left(\first{\frac{1}{3}}\right)\left(\first{-\frac{2}{3}}\right) \\ &\quad + \left(\second{\frac{2}{3}}\right)\left(\second{\frac{2}{3}}\right)+ \left(\third{\frac{2}{3}}\right)\third{c} = 0 \\ & \implies -\frac{2}{9} + \frac{4}{9} + \frac{2}{3}\third{c} = 0 \\ & \implies \third{c} = -\frac{1}{3} \end{aligned}
a\displaystyle \third{a} =23,b\displaystyle = \frac{2}{3}, \qquad \second{b} =13,c\displaystyle = \frac{1}{3}, \qquad \third{c} =−13\displaystyle = -\frac{1}{3}
5

Let QQ be an unknown 3×33 \times 3 orthogonal matrix, and let x=[12−2]\target{x} = \target{\begin{bmatrix} 1 \\ 2 \\ -2 \end{bmatrix}} and y=[212]y = \begin{bmatrix} 2 \\ 1 \\ 2 \end{bmatrix}. Without knowing the entries of QQ, find ∥Qx∥\|Q\target{x}\|, (Qx)⋅(Qy)(Q\target{x}) \cdot (Qy), and the angle between QxQ\target{x} and QyQy.

What happens to QTQQ^T Q inside a length or a dot product?

Because QQ keeps lengths the same, ∥Qx∥=∥x∥\|Q\target{x}\| = \|\target{x}\| (remember to take the square root at the end rather than stopping at ∥x∥2=9\|\target{x}\|^2 = 9):
∥Qx∥\displaystyle \|Q\target{x}\| =∥x∥\displaystyle = \|\target{x}\| =12+22+(−2)2\displaystyle = \sqrt{\target{1}^2 + \target{2}^2 + (\target{-2})^2} =9\displaystyle = \sqrt{9} =3\displaystyle = 3
Because QQ keeps dot products the same, (Qx)⋅(Qy)=x⋅y(Q\target{x}) \cdot (Qy) = \target{x} \cdot y:
(Qx)⋅(Qy)\displaystyle (Q\target{x}) \cdot (Qy) =(1)(2)+(2)(1)+(−2)(2)\displaystyle = (\target{1})(2) + (\target{2})(1) + (\target{-2})(2) =2+2−4\displaystyle = 2 + 2 - 4 =0\displaystyle = 0
A dot product of 00 means cos⁡θ=0\cos\theta = 0, so the vectors are perpendicular (don't confuse a dot product of 00 with an angle of 0∘0^\circ).
∥Qx∥\displaystyle \|Q\target{x}\| =3,(Qx)⋅(Qy)\displaystyle = 3, \qquad (Q\target{x}) \cdot (Qy) =0,angle\displaystyle = 0, \qquad \text{angle} =90∘\displaystyle = 90^\circ
6

Suppose Q1Q_1 and Q2Q_2 are n×nn \times n orthogonal matrices where Q1Q_1 is a rotation (det⁡Q1=+1\det Q_1 = +1) and Q2Q_2 is a reflection (det⁡Q2=−1\det Q_2 = -1). Show algebraically that M=Q1Q2TM = Q_1 Q_2^T is orthogonal, find M−1M^{-1} in terms of Q1Q_1 and Q2Q_2, and determine whether det⁡M\det M is +1+1 or −1-1.

When you transpose the product Q1Q2TQ_1 Q_2^T to compute MTMM^T M, what happens to the order of the two matrices?

Reverse the order of the matrices when transposing M=Q1Q2TM = Q_1 Q_2^T. Transposing in place without reversing the order gives the wrong expression Q1TQ2\fails{Q_1^T Q_2}, where the middle matrices cannot cancel:
MT\displaystyle M^T =(Q1Q2T)T\displaystyle = (Q_1 Q_2^T)^T =(Q2T)TQ1T\displaystyle = (Q_2^T)^T Q_1^T =Q2Q1T\displaystyle = Q_2 Q_1^T
Multiply MTMM^T M and use Q1TQ1=IQ_1^T Q_1 = I followed by Q2Q2T=IQ_2 Q_2^T = I:
MTM=(Q2Q1T)(Q1Q2T)=Q2(Q1TQ1)Q2T=Q2IQ2T=Q2Q2T=I\displaystyle \begin{aligned} M^T M &= (Q_2 Q_1^T)(Q_1 Q_2^T) \\ &= Q_2(Q_1^T Q_1)Q_2^T \\ &= Q_2 I Q_2^T = Q_2 Q_2^T = I \end{aligned}
This shows MM is orthogonal, so its inverse is M−1=MT=Q2Q1TM^{-1} = M^T = Q_2 Q_1^T. Finally, transposing Q2Q_2 does not change its determinant (det⁡Q2T=−1\det Q_2^T = -1, not +1+1), so following the reflection Q2TQ_2^T with the rotation Q1Q_1 leaves space flipped:
M−1\displaystyle M^{-1} =Q2Q1T,det⁡M\displaystyle = Q_2 Q_1^T, \qquad \det M =−1\displaystyle = -1