ISOMETRIES OF Rn

KEITH CONRAD

1. Introduction
An isometry of Rn is a function h : Rn → Rn that preserves the distance between vectors:

||h(v) − h(w)|| = ||v − w||

p
for all v and w in Rn , where ||(x1 , . . . , xn )|| = x21 + · · · + x2n .

Example 1.1. The identity transformation: id(v) = v for all v ∈ Rn .

Example 1.2. Negation: − id(v) = −v for all v ∈ Rn .

Example 1.3. Translation: fixing u ∈ Rn , let tu (v) = v + u. Easily ||tu (v) − tu (w)|| =
||v − w||.

Example 1.4. Rotations around points and reflections across lines in the plane are isome-
tries of R2 . Formulas for these isometries will be given in Example 3.3 and Section 4.

The effects of a translation, rotation (around the origin) and reflection across a line in
R2 are pictured below on sample line segments.

The composition of two isometries of Rn is an isometry. Is every isometry invertible? It

is clear that the three kinds of isometries pictured above (translations, rotations, reflections)
are each invertible (translate by the negative vector, rotate by the opposite angle, reflect a
second time across the same line).
In Section 2, we show the close link between isometries and the dot product on Rn ,
which is more convenient to use than distances due to its algebraic properties. Section 3
is about the matrices that act as isometries on on Rn , called orthogonal matrices. Section
4 describes the isometries of R and R2 geometrically. In Appendix A, we will look more
closely at reflections in Rn .
1
2 KEITH CONRAD

2. Isometries, dot products, and linearity

Using translations, we can reduce the study of isometries of Rn to the case of isometries
fixing 0.
Theorem 2.1. Every isometry of Rn can be uniquely written as the composition t ◦ k where
t is a translation and k is an isometry fixing the origin.
Proof. Let h : Rn → Rn be an isometry. If h = tw ◦ k, where tw is translation by a vector
w and k is an isometry fixing 0, then for all v in Rn we have h(v) = tw (k(v)) = k(v) + w.
Setting v = 0 we get w = h(0), so w is determined by h. Then k(v) = h(v)−w = h(v)−h(0),
so k is determined by h. Turning this around, if we define t(v) = v + h(0) and k(v) =
h(v)−h(0), then t is a translation, k is an isometry fixing 0, and h(v) = k(v)+h(0) = tw ◦k,
where w = h(0). 
Theorem 2.2. For a function h : Rn → Rn , the following are equivalent:
(1) h is an isometry and h(0) = 0,
(2) h preserves dot products: h(v) · h(w) = v · w for all v, w ∈ Rn .
Proof. The link between length and dot product is the formula
||v||2 = v · v.
Suppose h satisfies (1). Then for any vectors v and w in Rn ,
(2.1) ||h(v) − h(w)|| = ||v − w||.
As a special case, when w = 0 in (2.1) we get ||h(v)|| = ||v|| for all v ∈ Rn . Squaring both
sides of (2.1) and writing the result in terms of dot products makes it
(h(v) − h(w)) · (h(v) − h(w)) = (v − w) · (v − w).
Carrying out the multiplication,
(2.2) h(v) · h(v) − 2h(v) · h(w) + h(w) · h(w) = v · v − 2v · w + w · w.
The first term on the left side of (2.2) equals ||h(v)||2 = ||v||2 = v · v and the last term on
the left side of (2.2) equals ||h(w)||2 = ||w||2 = w · w. Canceling equal terms on both sides
of (2.2), we obtain −2h(v) · h(w) = −2v · w, so h(v) · h(w) = v · w.
Now assume h satisfies (2), so
(2.3) h(v) · h(w) = v · w
for all v and w in Rn . Therefore
||h(v) − h(w)||2 = (h(v) − h(w)) · (h(v) − h(w))
= h(v) · h(v) − 2h(v) · h(w) + h(w) · h(w)
= v · v − 2v · w + w · w by (2.3)
= (v − w) · (v − w)
= ||v − w||2 ,
so ||h(v) − h(w)|| = ||v − w||. Thus h is an isometry. Setting v = w = 0 in (2.3), we get
||h(0)||2 = 0, so h(0) = 0. 
Corollary 2.3. The only isometry of Rn fixing 0 and the standard basis is the identity.
 ISOMETRIES OF Rn 3

Proof. Let h : Rn → Rn be an isometry that satisfies

h(0) = 0, h(e1 ) = e1 , . . . , h(en ) = en .
Theorem 2.2 says
h(v) · h(w) = v · w
for all v and w in Rn . Fix v ∈ Rn and let w run over the standard basis vectors e1 , e2 , . . . , en ,
so we see
h(v) · h(ei ) = v · ei .
Since h fixes each ei ,
h(v) · ei = v · ei .
Writing v = c1 e1 + · · · + cn en , we get
h(v) · ei = ci
for all i, so h(v) = c1 e1 + · · · + cn en = v. As v was arbitrary, h is the identity on Rn . 

It is essential in Corollary 2.3 that the isometry fixes 0. An isometry of Rn fixing the
standard basis without fixing 0 need not be the identity! For example, reflection across the
line x + y = 1 in R2 is an isometry of R2 fixing (1, 0) and (0, 1) but not 0 = (0, 0). See
below.

Theorem 2.4. For a function h : Rn → Rn , the following are equivalent:

(1) h is an isometry and h(0) = 0,
(2) h is linear, and the matrix A such that h(v) = Av for all v ∈ Rn satisfies AA> = In .
Proof. Suppose h is an isometry and h(0) = 0. We want to prove linearity: h(v + w) =
h(v)+h(w) and h(cv) = ch(v) for all v and w in Rn and all c ∈ R. The mapping h preserves
dot products by Theorem 2.2:
h(v) · h(w) = v · w
for all v and w in Rn . For the standard basis e1 , . . . , en of Rn this says h(ei ) · h(ej ) =
ei · ej = δij , so h(e1 ), . . . , h(en ) is an orthonormal basis of Rn . Thus two vectors in Rn are
equal if they have the same dot product with each of h(e1 ), . . . , h(en ).
For all u in Rn we have
h(v + w) · h(u) = (v + w) · u
and
(h(v) + h(w)) · h(u) = h(v) · h(u) + h(w) · h(u) = v · u + w · u = (v + w) · u,
4 KEITH CONRAD

so h(v + w) · h(u) = (h(v) + h(w)) · h(u) for all u. Letting u = e1 , . . . , en shows h(v + w) =
h(v) + h(w). Similarly,
h(cv) · h(u) = (cv) · u = c(v · u) = c(h(v) · h(u)) = (ch(v)) · h(u),
so again letting u run through e1 , . . . , en tells us h(cv) = ch(v). Thus h is linear.
Let A be the matrix for h: h(v) = Av for all v ∈ Rn , where A has jth column h(ej ). We
want to show AA> = In . Since h preserves dot products, the condition h(v) · h(w) = v · w
for all v, w ∈ Rn says Av · Aw = v · w. The fundamental link between the dot product and
matrix transposes, which you should check, is that we can move a matrix to the other side
of a dot product by using its transpose:
(2.4) v · M w = M >v · w
for every n × n matrix M and v, w ∈ Rn . Using M = A and Av in place of v in (2.4),
Av · Aw = A> (Av) · w = (A> A)v · w.
This is equal to v · w for all v and w, so (A> A)v · w = v · w for all v and w in Rn . Since
the (i, j) entry of a matrix M is M ej · ei , letting v and w run through the standard basis of
Rn tells us A> A = In , so A is invertible. An invertible matrix commutes with its inverse,
so A> A = In ⇒ AA> = In .
For the converse, assume h(v) = Av for v ∈ Rn where AA> = In . Trivially h fixes 0. To
show h is an isometry, by Theorem 2.2 it suffices to show
(2.5) Av · Aw = v · w
for all v, w ∈ Rn . Since A and its inverse A> commute, we have A> A = In , so Av · Aw =
A> (Av) · w = (A> A)v · w = v · w. 
Corollary 2.5. Isometries of Rn are invertible, the inverse of an isometry is an isometry,
and two isometries on Rn that have the same values at 0 and any basis of Rn are equal.
This gives a second proof of Corollary 2.3 as a special case.
Proof. Let h : Rn → Rn be an isometry. By Theorem 2.1, h = k + h(0) where k is an
isometry of Rn fixing 0. Theorem 2.4 tells us there is an invertible matrix A such that
k(v) = Av for all v ∈ Rn , so
h(v) = Av + h(0).
−1 −1
This has inverse h (v) = A (v − h(0)). In particular, h is surjective.
The isometry condition ||h(v) − h(w)|| = ||v − w|| for all v and w in Rn implies ||v − w|| =
||h−1 (v) − h−1 (w)|| for all v and w in Rn by replacing v and w in the isometry condition
with h−1 (v) and h−1 (w). Thus h−1 is an isometry of Rn .
If h1 and h2 are isometries of Rn that are equal on 0 and a basis then the functions
k1 (v) = h1 (v) − h1 (0) and k2 (v) = h2 (v) − h2 (0) are linear and are equal on that basis, so
by linearity k1 = k2 on Rn . That is, h1 (v) − h1 (0) = h2 (v) − h2 (0) for all v in Rn . Since
h1 (0) = h2 (0) we get h1 = h2 on Rn . 
Remark 2.6. That isometries of Rn fixing 0 are linear and invertible is a special case of
the following more general result: for a finite-dimensional vector space V over an arbitrary
field and a nondegenerate bilinear form B on V , a function A : V → V for which B(v, w) =
B(A(v), A(w)) for all v and w in V must be linear and invertible. A more general version
of this is due to A. Vogt [3, Lemma 1.5, Theorem 2.4], and a proof can be found there
or in my answer at https://math.stackexchange.com/questions/137139. A physically
 ISOMETRIES OF Rn 5

interesting example of this over R besides Rn with its usual dot product is 4-dimensional
space (x, y, z, ct) with the indefinite bilinear form associated to x2 + y 2 + z 2 − c2 t2 in special
relativity (Minkowski spacetime).
Definition 2.7. In Rn , a set of n + 1 points P0 , P1 , . . . , Pn is said to be in general position
if they don’t all lie in a hyperplane.
This concept abstracts the idea of 3 points in R2 not being collinear. In the definition,
the hyperplanes in Rn are translated subspaces of dimension n − 1, so they need not pass
through the origin. For example, a line in R2 need not be a linear subspace of R2 since a
line doesn’t have to contain the origin. Three points in R2 are in general position if no line
passes through all of them and four points in R3 are in general position if no plane passes
through all of them. Saying P0 , P1 , . . . , Pn are in general position in Rn does not mean these
n + 1 points are linearly independent as vectors in Rn , but rather that the n differences
P1 − P0 , . . . , Pn − P0 are linearly independent vectors: a nontrivial linear relation would
place these n differences, along with 0, in a common subspace of dimension n − 1, so adding
P0 to all of the differences and to 0 would put P0 , P1 , . . . , Pn in a common hyperplane.
Adding a common vector to points in general position keeps them in general position
since the added vector cancels out when taking differences.
Corollary 2.8. Let P0 , P1 , . . . , Pn be n + 1 points in Rn in “general position”. Two isome-
tries of Rn that are equal at P0 , . . . , Pn are the same.
Proof. We know isometries of Rn are invertible. If h1 and h2 are isometries of Rn with the
same values at each Pi then h−12 ◦ h1 is an isometry that fixes each Pi . Therefore to prove
h1 = h2 it suffices to show an isometry of Rn that fixes P0 , . . . , Pn is the identity.
Let h be an isometry of Rn such that h(Pi ) = Pi for 0 ≤ i ≤ n. Set t(v) = v − P0 , which
is a translation. Then tht−1 is an isometry with formula
(tht−1 )(v) = h(v + P0 ) − P0 .
Thus (tht−1 )(0) = h(P0 )−P0 = 0, so tht−1 is linear by Theorem 2.4. Also (tht−1 )(Pi −P0 ) =
h(Pi ) − P0 = Pi − P0 .
Since P0 , . . . , Pn are in general position, the differences P1 − P0 , . . . , Pn − P0 form a basis
of Rn . Therefore by Corollary 2.5, tht−1 is the identity, so h is the identity. 

3. Orthogonal matrices
We have seen that the isometries of Rn that fix 0 come from matrices A such that
AA> = In . These matrices have a name.
Definition 3.1. An n × n matrix A is called orthogonal if AA> = In , or equivalently if
A> A = In .
A matrix is orthogonal when its transpose is its inverse. Since det(A> ) = det A, an
orthogonal matrix A satisfies (det A)2 = 1, so det A = ±1. (For n ≥ 2 not all matrices with
determinant ±1 are orthogonal, such as ( 35 12 ). The orthogonal 1 × 1 matrices are ±1.)
Example 3.2. Negation on Rn (Example 1.2) is an isometry that is described by the
matrix −In , which is orthogonal: (−In )(−In )> = (−In )(−In ) = In .
Example 3.3. Let n = 2. By algebra, AA> = I2 if and only if A = ( ab −εb εa
), where
2 2 −
a +b = 1 and ε = ±1. Writing a = cos θ and b = sin θ, we get the matrices ( sin θ cos θθ ) and
cos θ sin
6 KEITH CONRAD

( cos θ sin θ
sin θ − cos θ ). Algebraically, these types of matrices are distinguished by their determinants:
the first type has determinant 1 and the second type has determinant −1.
The geometric effects of these two types of matrices differ. Below on the left, ( cos θ − sin θ )
sin θ cos θ
is a counterclockwise rotation by angle θ around the origin. Below on the right, ( cos θ sin θ
sin θ − cos θ )
is a reflection across the line through the origin at angle θ/2 with respect to the positive
x-axis. (Check ( cos θ sin θ
sin θ − cos θ ) squares to the identity, as any reflection should.)

A = ( cos θ − sin θ ) A = ( cos θ sin θ

sin θ cos θ A(v) sin θ − cos θ )
v v
w w
A(w)

A(v)
A(w)

Let’s explain why ( cos θ sin θ

sin θ − cos θ ) is a reflection at angle θ/2. See the figure below. Pick a
line L through the origin, say at an angle ϕ with respect to the positive x-axis. To find a
formula for reflection across L, we’ll use a basis of R2 with
one vector on L and the other
vector perpendicular to L. The unit vector u1 = cos ϕ
sin ϕ lies on L and the unit vector
u2 = − cos
sin ϕ 2


ϕ is perpendicular to L. For any v ∈ R , write v = c1 u1 + c2 u2 with c1 , c2 ∈ R.

v
u2 L
c2 u2
c1 u1
ϕ u1
s(v)
−c2 u2

The reflection of v across L is s(v) = c1 u1 − c2 u2 . Writing a = cos ϕ and b = sin ϕ (so

a2 + b2 = 1), in standard coordinates this becomes

        
a −b c1 a − c2 b a −b c1
(3.1) v = c1 u1 + c2 u2 = c1 + c2 = =
b a c1 b + c2 a b a c2
 ISOMETRIES OF Rn 7

and in a similar way

s(v) = c1 u1 − c2 u2
  
a b c1
=
b −a c2
  −1
a b a −b
= v by (3.1)
b −a b a
  
a b a b
= v
b −a −b a
 2
a − b2

2ab
= v.
2ab −(a2 − b2 )
cos(2ϕ) sin(2ϕ)
By the sine and cosine duplication formulas, the last matrix is ( sin(2ϕ) − cos(2ϕ)
). Therefore
( cos θ sin θ
sin θ − cos θ ) is a reflection across the line through the origin at angle θ/2.

We return to orthogonal n × n matrices for any n ≥ 1. The geometric meaning of the

condition A> A = In is that the columns of A are mutually perpendicular unit vectors
(check!). From this we see how to create orthogonal matrices: starting with an orthonormal
basis of Rn , an n × n matrix having this basis as its columns (in any order) is an orthogonal
matrix, and all n × n orthogonal matrices arise in this way.
Let On (R) denote the set of n × n orthogonal matrices:
(3.2) On (R) = {A ∈ GLn (R) : AA> = In }.
Theorem 3.4. The set On (R) is a group under matrix multiplication.
Proof. Clearly In ∈ On (R). If A and B are in On (R), then
(AB)(AB)> = ABB > A> = AA> = In ,
so AB ∈ On (R). For A ∈ On (R), we have A−1 = A> and
(A−1 )(A−1 )> = A> (A> )> = A> A = In .
Therefore A−1 ∈ On (R). 
The link between isometries and dot products (Theorem 2.2) gives us a more geometric
description of On (R) than (3.2):
(3.3) On (R) = {A ∈ GLn (R) : Av · Aw = v · w for all v, w ∈ Rn }.
The label “orthogonal matrix” is very unfortunate. It suggests that such matrices should
be the ones that preserve orthogonality of vectors:
(3.4) v · w = 0 =⇒ Av · Aw = 0
for all v and w in Rn . While orthogonal matrices do satisfy (3.4), since (3.4) is a special
case of the condition Av · Aw = v · w in (3.3), many matrices satisfy (3.4) and are not
orthogonal matrices! That is, orthogonal matrices (which, by definition, preserve all dot
products) are not the only matrices that preserve orthogonality of vectors (dot products
equal to 0). A simple example of a nonorthogonal matrix satisfying (3.4) is a scalar matrix
cIn , where c 6= ±1. Since (cv) · (cw) = c2 (v · w), cIn does not preserve dot products in
general but it does preserve dot products equal to 0. It’s natural to ask which matrices
8 KEITH CONRAD

besides orthogonal matrices preserve orthogonality. Here is the complete answer, which
shows they are not that far from being orthogonal.
Theorem 3.5. An n × n real matrix A satisfies (3.4) if and only if A is a scalar multiple
of an orthogonal matrix.
Proof. If A = cA0 where A0 is orthogonal, then Av · Aw = c2 (A0 v · A0 w) = c2 (v · w), so if
v · w = 0 then Av · Aw = 0.
Now assume A satisfies (3.4). Then the vectors Ae1 , . . . , Aen are mutually perpendicular,
so the columns of A are perpendicular to each other. We want to show that they have the
same length.
Note that ei + ej ⊥ ei − ej when i 6= j, so by (3.4) and linearity Aei + Aej ⊥ Aei − Aej .
Writing this in the form (Aei + Aej ) · (Aei − Aej ) = 0 and expanding, we are left with
Aei · Aei = Aej · Aej , so ||Aei || = ||Aej ||. Therefore the columns of A are mutually
perpendicular vectors with the same length. Call this common length c. If c = 0 then
A = O = 0 · In . If c 6= 0 then the matrix (1/c)A has an orthonormal basis as its columns, so
it is an orthogonal matrix. Therefore A = c((1/c)A) is a scalar multiple of an orthogonal
matrix. 
Since a composition of isometries is an isometry and isometries are invertible with the
inverse of an isometry being an isometry, isometries form a group under composition. We
will describe the elements of this group and show how the group law looks in that description.
Theorem 3.6. For A ∈ On (R) and w ∈ Rn , the function hA,w : Rn → Rn given by
hA,w (v) = Av + w = (tw A)(v)
is an isometry. Moreover, every isometry of Rn has this form for unique A and w.
Proof. The indicated formula always gives an isometry, since it is the composition of a
translation and orthogonal matrix transformation, which are both isometries.
To show every isometry of Rn has the form hA,w for some A and w, let h : Rn → Rn be
an isometry. By Theorem 2.1, h = k + h(0) where k is an isometry of Rn fixing 0. Theorem
2.4 tells us there is an A ∈ On (R) such that k(v) = Av for all v ∈ Rn , so
h(v) = k(v) + h(0) = Av + h(0) = hA,w (v),
where w = h(0).
If hA,w = hA0 ,w0 as functions on Rn , then evaluating both sides at 0 gives w = w0 .
Therefore Av + w = A0 v + w for all v, so Av = A0 v for all v, which implies A = A0 . 
Let Iso(Rn ) denote the group of isometries of Rn . Its elements have the form hA,w by
Theorem 3.6. Here is what composition of such mappings looks like:
hA,w (hA0 ,w0 (v)) = A(A0 v + w0 ) + w
= AA0 v + Aw0 + w
= hAA0 ,Aw0 +w (v).
This is similar to the multiplication law in the ax + b group:
 0 0  
aa0 ab0 + b
 
a b a b
= .
0 1 0 1 0 1
In fact, if we write an isometry hA,w ∈ Iso(Rn ) as an (n + 1) × (n + 1) matrix ( A0 w1 ), where
the 0 in the bottom is a row vector of n zeros, then the composition law in Iso(Rn ) is
 ISOMETRIES OF Rn 9

multiplication of the corresponding (n + 1) × (n + 1) matrices, so
Iso(Rn ) can be viewed as a

subgroup of GLn+1 (R), acting on Rn as the column vectors v1 in Rn+1 (not a subspace!).

4. Geometric description of isometries of R and R2

Let’s classify the isometries of Rn for n = 1 and n = 2.
Since O1 (R) = {±1}, the isometries of R are the functions h(x) = x+c and h(x) = −x+c
for c ∈ R. (Of course, this case can be worked out easily without the earlier material.)
Now consider isometries of R2 . Write an isometry h ∈ Iso(R2 ) as h(v) = Av + w with
A ∈ O2 (R) and w ∈ R2 . By Example 3.3, A is a rotation or reflection, depending on det A.
There turn out to be four possibilities for h: translations, rotations, reflections, and glide
reflections. A glide reflection is the composition of a reflection and a nonzero translation in
a direction parallel to the line of reflection. A picture of a glide reflection is in the figure
below, where the (horizontal) line of reflection is dashed and the translation is a movement
to the right.

The image above, which includes “before” and “after” states, suggests a physical inter-
pretation of a glide reflection: it is the result of turning the plane in space like a half-turn
of a screw. A more picturesque image, suggested to me by Michiel Vermeulen, is the effect
of successive steps with a left foot and then a right foot in the sand or snow (if your feet
are mirror reflections).
The possibilities for isometries of f are collected in Table 1 below. It describes how the
type of an isometry h is determined by det A and the geometry of the set of fixed points of h
(solutions to h(v) = v): empty, a point, a line, or the plane. (The only isometry belonging
to more than one of the four possibilities is the identity, which is both a translation and a
rotation, so we make the identity its own row in the table.) The table also shows how a
description of the fixed points can be obtained algebraically from A and w.

Isometry Condition Fixed pts

Identity A = I2 , w = 0 R2
Nonzero Translation A = I2 , w 6= 0 ∅
Nonzero Rotation det A = 1, A 6= I2 (I2 − A)−1 w
Reflection det A = −1, Aw = −w w/2 + ker(A − I2 )
Glide Reflection det A = −1, Aw 6= −w ∅
2
Table 1. Isometries of R : h(v) = Av + w, A ∈ O2 (R).

To justify the information in the table we move down the middle column. The first two
rows are obvious, so we start with the third row.
10 KEITH CONRAD

Row 3: Suppose det A = 1 and A 6= I2 , so A = ( cos θ − sin θ ) for some θ and cos θ 6= 1. We
sin θ cos θ
want to show h is a rotation. First of all, h has a unique fixed point: v = Av + w precisely
when w = (I2 − A)v. We have det(I2 − A) = 2(1 − cos θ) 6= 0, so I2 − A is invertible and
p = (I2 − A)−1 w is the fixed point of h. Then w = (I2 − A)p = p − Ap, so
(4.1) h(v) = Av + (p − Ap) = A(v − p) + p.
Since A is a rotation by θ around the origin, (4.1) shows h is a rotation by θ around P .
Rows 4, 5: Suppose det A = −1, so A = ( cos θ sin θ 2
sin θ − cos θ ) for some θ and A = I2 . We again
look at fixed points of h. As before, h(v) = v for some v if and only if w = (I2 − A)v.
But unlike the previous case, now det(I2 − A) = 0 (check!), so I2 − A is not invertible and
therefore w may or may not be in the image of I2 − A. When w is in the image of I2 − A,
we will see that h is a reflection. When w is not in the image of I2 − A, we will see that h
is a glide reflection.
Suppose the isometry h(v) = Av + w with det A = −1 has a fixed point. Then w/2 must
be a fixed point. Indeed, let p be any fixed point, so p = Ap + w. Since A2 = I2 ,
Aw = A(p − Ap) = Ap − p = −w,
so w w 1 w
h =A + w = Aw + w = .
2 2 2 2
Conversely, if h(w/2) = w/2 then A(w/2) + w = w/2,, so Aw = −w.
Thus h has a fixed point if and only if Aw = −w, in which case
 w w
(4.2) h(v) = Av + w = A v − + .
2 2
Since A is a reflection across some line L through 0, (4.2) says h is a reflection across the
parallel line w/2 + L passing through w/2. See the figure below. (Algebraically, we can say
L = {v : Av = v} = ker(A − I2 ). Since A − I2 is not invertible and not identically 0, its
kernel really is 1-dimensional.)

v
Av

L w/2
w/2 + L h(v)
w
Now assume h has no fixed point, so Aw 6= −w. We will show h is a glide reflection. (The
formula h = Av +w shows h is the composition of a reflection and a nonzero translation, but
w need not be parallel to the line of reflection of A, which is ker(A − I2 ), so this formula for
h does not show directly that h is a glide reflection.) We will now take stronger advantage
of the fact that A2 = I2 .
Since O = A2 − I2 = (A − I2 )(A + I2 ) and A 6= ±I2 (after all, det A = −1), A + I2 and
A − I2 are not invertible. Therefore the subspaces
W1 = ker(A − I2 ), W2 = ker(A + I2 )
 ISOMETRIES OF Rn 11

are both nonzero, and neither is the whole plane, so W1 and W2 are both one-dimensional.
We already noted that W1 is the line of reflection of A (fixed points of A form the kernel of
A − I2 ). It turns out that W2 is the line perpendicular to W1 . To see why, pick w1 ∈ W1
and w2 ∈ W2 , so
Aw1 = w1 , Aw2 = −w2 .
Then, since Aw1 · Aw2 = w1 · w2 by orthogonality of A, we have
w1 · (−w2 ) = w1 · w2 .
Thus w1 · w2 = 0, so w1 ⊥ w2 .
Now we are ready to show h is a glide reflection. Pick nonzero vectors wi ∈ Wi for i = 1, 2,
and use {w1 , w2 } as a basis of R2 . Write w = h(0) in terms of this basis: w = c1 w1 + c2 w2 .
To say there are no fixed points for h is the same as Aw 6= −w, so w 6∈ W2 . That is, c1 6= 0.
Then
(4.3) h(v) = Av + w = (Av + c2 w2 ) + c1 w1 .
Since A(c2 w2 ) = −c2 w2 , our previous discussion shows v 7→ Av + c2 w2 is a reflection
across the line c2 w2 /2 + W1 . Since c1 w1 is a nonzero vector in W1 , (4.3) exhibits h as the
composition of a reflection across the line c2 w2 /2 + W1 and a nonzero translation by c1 w1 ,
whose direction is parallel to the line of reflection, so h is a glide reflection.
We have now justified the information in Table 1. Each row describes a different kind
of isometry. Using fixed points it is easy to distinguish the first four rows from each other
and to distinguish glide reflections from any isometry besides translations. A glide reflection
can’t be a translation since any isometry of R2 is uniquely of the form hA,w , and translations
have A = I2 while glide reflections have det A = −1.
Lemma 4.1. A composition of two reflections of R2 is a translation or a rotation.
Proof. The product of two matrices with determinant −1 has determinant 1, so the com-
position of two reflections has the form v 7→ Av + w where det A = 1. Such isometries
are translations or rotations by Table 1 (consider the identity to be a trivial translation or
rotation). 
In Example A.2 we will express any translation in Rn as the composition of two reflec-
tions.
Theorem 4.2. Each isometry of R2 is a composition of at most 2 reflections except for
glide reflections, which are a composition of 3 (and no fewer) reflections.
Proof. We check the theorem for each type of isometry in Table 1 besides reflections, for
which the theorem is obvious.
The identity is the square of any reflection.
For a translation t(v) = v + w, let A be the matrix representing the reflection across
the line w⊥ . Then Aw = −w. Set s1 (v) = Av + w and s2 (v) = Av. Both s1 and s2 are
reflections, and (s1 ◦ s2 )(v) = A(Av) + w = v + w since A2 = I2 .
Now consider a rotation, say h(v) = A(v − p) + p for some A ∈ O2 (R) with det A = 1
and p ∈ R2 . We have h = t ◦ r ◦ t−1 , where t is translation by p and r(v) = Av is a rotation
around the origin. Let A0 be any reflection matrix (e.g., A0 = ( 10 −10 )). Set s (v) = AA0 v
1
0
and s2 (v) = A v. Both s1 and s2 are reflections and r = s1 ◦ s2 (check). Therefore
(4.4) h = t ◦ r ◦ t−1 = (t ◦ s1 ◦ t−1 ) ◦ (t ◦ s2 ◦ t−1 ).
12 KEITH CONRAD

The conjugate of a reflection by a translation (or by any isometry, for that matter) is another
reflection, as an explicit calculation using Table 1 shows. Thus, (4.4) expresses the rotation
h as a composition of 2 reflections.
Finally we consider glide reflections. Since this is the composition of a translation and
a reflection, it is a composition of 3 reflections. We can’t use fewer reflections to get a
glide reflection, since a composition of two reflections is either a translation or a rotation
by Lemma 4.1 and we know that a glide reflection is not a translation or rotation (or
reflection). 
In Table 2 we record the minimal number of reflections whose composition can equal a
particular type of isometry of R2 .

Isometry Min. Num. Reflections dim(fixed set)

Identity 0 2
Nonzero Translation 2 0
Nonzero Rotation 2 0
Reflection 1 1
Glide Reflection 3 0
Table 2. Counting Reflections in an Isometry

That each isometry of R2 is a composition of at most 3 reflections can be proved geomet-

rically, without recourse to a prior classification of all isometries of the plane. We will give
a rough sketch of the argument. We will take for granted (!) that an isometry that fixes at
least two points is a reflection across the line through those points or is the identity. (This
is related to Corollary 2.3 when n = 2.) Pick any isometry h of R2 . We may suppose h
is not a reflection or the identity (the identity is the square of any reflection), so h has at
most one fixed point. If h has one fixed point, say P , choose Q 6= P . Then h(Q) 6= Q and
the points Q and h(Q) lie on a common circle centered at P (because h(P ) = P ). Let s be
the reflection across the line through P that is perpendicular to the line connecting Q and
h(Q). Then s ◦ h fixes P and Q, so s ◦ h is the identity or is a reflection. Thus h = s ◦ (s ◦ h)
is a reflection or a composition of two reflections. If h has no fixed points, pick any point P .
Let s be the reflection across the perpendicular bisector of the line connecting P and h(P ),
so s ◦ h fixes P . Thus s ◦ h has a fixed point, so our previous argument shows s ◦ h is either
the identity, a reflection, or the composition of two reflections, so h is the composition of
at most 3 reflections.
A byproduct of this argument, which did not use the classification of isometries, is another
proof that all isometries of R2 are invertible: any isometry is a composition of reflections
and reflections are invertible.
From the fact that all isometries fixing 0 in R and R2 are rotations or reflections, the
following general description can be proved about isometries of any Euclidean space in terms
of rotations and reflections on one-dimensional and two-dimensional subspaces.
Theorem 4.3. If h is an isometry of Rn that fixes 0 then there is an orthogonal decompo-
sition Rn = W1 ⊕ W2 ⊕ · · · ⊕ Wm such that dim(Wi ) = 1 or 2 for all i, and the restriction
of h to Wi is a rotation unless i = m and dim(Wm ) = 1 and det h = −1, in which case the
restriction of h to Wm is a reflection.
Proof. See [1, Theorem 6.47] or [2, Cor. to Theorem 2]. 
 ISOMETRIES OF Rn 13

Appendix A. Reflections
A reflection is an isometry of Rn that fixes all the points in a chosen hyperplane and
interchanges the position of points along each line perpendicular to that hyperplane at
equal distance from it. These isometries play a role that is analogous to transpositions in
the symmetric group. Reflections, like transpositions, have order 2.
Let’s look first at reflections across hyperplanes that contain the origin. Let H be a
hyperplane containing the origin through which we wish to reflect. Set L = H ⊥ , so L is a
one-dimensional subspace. Every v ∈ Rn can be written uniquely in the form v = w + u,
where w ∈ H and u ∈ L. The reflection across H, by definition, is the function
(A.1) s(v) = s(w + u) = w − u.
That is, s fixes H = u⊥ and acts like −1 on L = Ru. From the formula defining s, it is
linear in v. Since w ⊥ u, ||s(v)|| = ||w|| + ||u|| = ||v||, so by linearity s is an isometry:
||s(v) − s(w)|| = ||s(v − w)|| = ||v − w||.
Since s is linear, it can be represented by a matrix. To write this matrix simply, pick an
orthogonal basis {v1 , . . . , vn−1 } of H and let vn be a nonzero vector in L = H ⊥ , so vn is
orthogonal to H. Then
s(c1 v1 + · · · + cn vn ) = c1 v1 + · · · + cn−1 vn−1 − cn vn .
The matrix for s has 1’s along the diagonal except for −1 in the last position:
    
c1 1 ··· 0 0 c1
 ..   .. . . .. ..   .. 
(A.2)
 .

  .
= . . . 
 .

.

 cn−1   0 · · · 1 0   cn−1 
−cn 0 · · · 0 −1 cn
The matrix in (A.2) represents s relative to a convenient choice of basis. In particular, from
the matrix representation we see det s = −1: every reflection in On (R) has determinant
−1. Notice the analogy with transpositions in the symmetric group, which have sign −1.
We now derive another formula for s, which will look more complicated than what we
have seen so far but should be considered more fundamental. Fix a nonzero vector u on the
line L = H ⊥ . Since Rn = H ⊕ L, any v ∈ Rn can be written as w + cu, where w ∈ H and
c ∈ R. Since w ⊥ L, v · u = c(u · u), so c = (v · u)/(u · u). Then
v·u
(A.3) s(v) = w − cu = v − 2cu = v − 2 u.
u·u
The last expression is our desired formula for s(v). Note for all v that s(v) · u = −v · u.
It is standard to label the reflection across a hyperplane containing the origin using
a vector in the orthogonal complement to the hyperplane, so we write s in (A.3) as su .
This is the reflection in the hyperplane u⊥ , so su (u) = −u. By (A.3), sau = su for any
a ∈ R − {0}, which makes geometric sense since (au)⊥ = u⊥ , so the reflection in the
hyperplane orthogonal to u and to au is the same. Moreover, H is the set of points fixed
by su , and we can confirm this with (A.3): su (v) = 0 if and only if v · u = 0, which means
v ∈ u⊥ = H.
To get a formula for the reflection across any hyperplane in Rn (not just those containing
the origin), we use the following lemma to describe any hyperplane.
14 KEITH CONRAD

Lemma A.1. Every hyperplane in Rn has the form Hu,c = {v ∈ Rn : v · u = c} for some
nonzero u ∈ Rn that is orthogonal to the hyperplane and some c ∈ R. The hyperplane
contains 0 if and only if c = 0.
Proof. Let H be a hyperplane and choose w ∈ H. Then H − w is a hyperplane containing
the origin. Fix a nonzero vector u that is perpendicular to H. Since H − w is a hyperplane
through the origin parallel to H, a vector v lies in H if and only if v − w ⊥ u, which is
equivalent to v · u = w · u. Thus H = Hu,c for c = w · u. 

Below are hyperplanes (lines) in R2 of the form H(2,1),c = {v : v · (2, 1) = c}.

(2, 1)

w1

w2

c=0c=2c=4
As the figure suggests, the different hyperplanes Hu,c as c varies are parallel to each other.
Specifically, if w ∈ Hu,c then Hu,c = Hu,0 + w (check!). (The choice of w in Hu,c affects how
Hu,0 is translated over to Hu,c , since adding w to Hu,0 sends 0 to w. Compare in the above
figure how Hu,0 is carried onto Hu,4 using translation by w1 and by w2 .)
In the family of parallel hyperplanes {Hu,c : c ∈ R}, we can replace u with any nonzero
scalar multiple, since Hau,c = Hu,c/a , so {Hu,c : c ∈ R} = {Hau,c : c ∈ R}. Geometrically
this makes sense, since the importance of u relative to the hyperplanes is that it is an
orthogonal direction, and au also provides an orthogonal direction to the same hyperplanes.
To reflect points across a hyperplane H, fix a nonzero vector w ∈ H. Geometric intuition
suggests that to reflect across H we can subtract w, then reflect across H − w (a hyperplane
through the origin), and then add w back. In the figure below, this corresponds to moving
from P to Q (subtract w from P ) to Q0 (reflect Q across H − w) to P 0 (add w to Q0 ),
getting the reflection of P across H.

Q0

w P0
Q
P
H −w H
Therefore reflection across H should be given by the formula
(A.4) s0 (v) = s(v − w) + w,
 ISOMETRIES OF Rn 15

where s is reflection across H − w. Setting H = Hu,c by Lemma A.1, where u is a nonzero

vector orthogonal to H, c = u · w (since w ∈ H) and by (A.3) and (A.4)
 
0 (v − w) · u v·u−c
(A.5) s (v) = (v − w) − 2 u+w =v−2 u.
u·u u·u
The following properties show (A.5) is the reflection across the hyperplane Hu,c .
• If v ∈ Hu,c then v · u = c, so (A.5) implies s0 (v) = v: s0 fixes points in Hu,c .
• For any v in Rn , the average 12 (v + s0 (v)), which is the
midpoint of the segment
0 v·u−c
connecting v and s (v), lies in Hu,c : it equals v − u·u u, whose dot product with
u is c.
• For any v in Rn the difference v − s0 (v), which is the direction of the segment
connecting v and s0 (v), is perpendicular to Hu,c since, by (A.5), it lies in Ru = H ⊥ .

Example A.2. We use (A.5) to show any nonzero translation tu (v) = v + u is the compo-
sition of two reflections. Set H = u⊥ = Hu,0 and write su for the reflection across H and
s0u for the reflection across H + u, the hyperplane parallel to H that contains u. By (A.3)
and (A.5),
   
0 su (v) · u − u · u −v · u
su (su (v)) = su (v) − 2 u = su (v) − 2 − 1 u = v + 2u,
u·u u·u

so s0u ◦ su = t2u . This is true for all u, so tu = s0u/2 ◦ su/2 .

These formulas show any translation is a composition of two reflections across hyperplanes
perpendicular to the direction of the translation.

The figure below illustrates Example A.2 in the plane, with u being a vector along the
x-axis. Reflecting v and w across H = u⊥ and then across H + u is the same as translation
of v and w by 2u.

H H +u
s(v) v s0 (s(v))

u 2u

w s(w) s0 (s(w))

Theorem A.3. Let w and w0 be distinct in Rn . There is a unique reflection s in Rn such

that s(w) = w0 . This reflection is in On (R) if and only if w and w0 have the same length.

Proof. A reflection taking w to w0 has a fixed hyperplane that contains the average 21 (w+w0 )
and is orthogonal to w − w0 . Therefore the fixed hyperplane of a reflection taking w to w0
must be Hw−w0 ,c for some c. Since 21 (w + w0 ) ∈ Hw−w0 ,c , we have c = (w − w0 ) · 12 (w + w0 ) =
1 0 0 0
2 (w · w − w · w ). Thus the only reflection that could send w to w is the one across the
hyperplane Hw−w0 , 1 (w·w−w0 ·w0 ) .
2
16 KEITH CONRAD

Let’s check that reflection across this hyperplane does send w to w0 . Its formula, by
(A.5), is
v · (w − w0 ) − c
 
s(v) = v − 2 (w − w0 ),
(w − w0 ) · (w − w0 )
where c = 12 (w · w − w0 · w0 ). When v = w, the coefficient of w − w0 in the above formula
becomes −1, so s(w) = w − (w − w0 ) = w0 .
If w and w0 have the same length then w · w = w0 · w0 , so c = 0 and that means s has
fixed hyperplane Hw−w0 ,0 . Therefore s is a reflection fixing 0, so s ∈ On (R). Conversely, if
s ∈ On (R) then s(0) = 0, which implies 0 ∈ Hw−w0 ,c , so c = 0, and therefore w · w = w0 · w0 ,
which means w and w0 have the same length.
To illustrate techniques, when w and w0 are distinct vectors in Rn with the same length
let’s construct a reflection across a hyperplane through the origin that sends w to w0 geo-
metrically, without using the algebraic formulas for reflections and hyperplanes.
If w and w0 are on the same line through the origin then w0 = −w (the only vectors on
Rw with the same length as w are w and −w). For the reflection s across the hyperplane
w⊥ , s(w) = −w = w0 .
If w and w0 are not on the same line through the origin then the span of w and w0 is a
plane. The vector v = w + w0 is nonzero and lies on the line in this plane that bisects the
angle between w and w0 . (See the figure below.) Let u be a vector in this plane orthogonal
to v, so writing w = av + bu we have w0 = av − bu.1 Letting s be the reflection in Rn across
the hyperplane u⊥ , which contains Rv (and contains more than Rv when n > 2), we have
s(v) = v and s(u) = −u, so s(w) = s(av + bu) = av − bu = w0 .

v
w
w0
u


We have already noted that reflections in On (R) are analogous to transpositions in the
symmetric group Sn : they have order 2 and determinant −1, just as transpositions have
order 2 and sign −1. The next theorem, due to E. Cartan, is the analogue for On (R) of the
generation of Sn by transpositions.
Theorem A.4 (Cartan). The group On (R) is generated by its reflections.
Note that a reflection in On (R) fixes 0 and therefore its fixed hyperplane contains the
origin, since a reflection does not fix any point outside its fixed hyperplane.

1This is geometrically clear, but algebraically tedious. Since v = w + w 0 , we have w 0 = v − w = (1 − a)v − bu,
so to show w0 = av − bu we will show a = 12 . Since v ⊥ u, w · v = a(v · v). The vectors w and w0 have the
same length, so w · v = w · (w + w0 ) = w · w + w · w0 and v · v = (w + w0 ) · (w + w0 ) = 2(w · w + w · w0 ), so
w · v = 12 (v · v). Comparing this with w · v = a(v · v), we have a = 21 .
 ISOMETRIES OF Rn 17

Proof. We argue by induction on n. The theorem is trivial when n = 1, since O1 (R) = {±1}.
Let n ≥ 2. (While the case n = 2 was treated in Theorem 4.2, we will reprove it here.)
Pick h ∈ On (R), so h(en ) and en have the same length. If h(en ) 6= en , by Theorem A.3
there is a (unique) reflection s in On (R) such that s(h(en )) = en , so the composite isometry
sh = s ◦ h fixes en . If h(en ) = en then we can write s(h(en )) = en where s is the identity
on Rn . We will use s with this meaning (reflection or identity) below.
Any element of On (R) preserves orthogonality, so sh sends the hyperplane H := e⊥ n =
R n−1 ⊕ {0} back to itself and is the identity on the line Ren . Since e⊥ n = R n−1 ⊕ {0} has

dimension n − 1, by induction2 there are a finite number of reflections s1 , . . . , sm in H fixing

the origin such that
sh|H = s1 s2 · · · sm .
Any reflection H → H that fixes 0 extends naturally to a reflection of Rn fixing 0, by
declaring it to be the identity on the line H ⊥ = Ren and extending by linearity from the
behavior on H and H ⊥ .3 Write si for the extension of si to a reflection on Rn in this way.
Consider now the two isometries
sh, s1 s2 · · · sm
of Rn . They agree on H = e⊥
n and they each fix en . Thus, by linearity, we have equality as
functions on Rn :
sh = s1 s2 · · · sm .
Therefore h = s−1 s1 s2 · · · sm . 
From the proof, if (sh)|H is a composition of m isometries of H fixing 0 that are the
identity or reflections then h is a composition of m + 1 isometries of Rn fixing 0 that are
the identity or reflections. Therefore every element of On (R) is a composition of at most n
elements of On (R) that are the identity or reflections (in other words, from m ≤ n−1 we get
m + 1 ≤ n). If h is not the identity then such a decomposition of h must include reflections,
so by removing the identity factors we see h is a composition of at most n reflections.
The identity on Rn is a composition of 2 reflections. This establishes the stronger form of
Cartan’s theorem: every element of On (R) is a composition of at most n reflections (except
for the identity when n = 1, unless we use the convention that the identity is a composition
of 0 reflections).
Remark A.5. Cartan’s theorem can be deduced from the decomposition of Rn in Theorem
4.3. Let a be the number of 2-dimensional Wi ’s and b be the number of 1-dimensional Wi ’s,
so 2a + b = n and h acts as a rotation on any 2-dimensional Wi . By Theorem 4.2, any
rotation of Wi is a composition of two reflections in Wi . A reflection in Wi can be extended
to a reflection in Rn by setting it to be the identity on the other Wj ’s. If Wi is 1-dimensional
then h is the identity on Wi except perhaps once, in which case b ≥ 1 and h is a reflection
on that Wi . Putting all of these reflections together, we can express h as a composition of
at most 2a reflections if b = 0 and at most 2a + 1 reflections if b ≥ 1. Either way, h is a
2Strictly speaking, since H is not Rn−1 , to use induction we really should be proving the theorem not
just for orthogonal transformations of the Euclidean spaces Rn , but for orthogonal transformations of their
subspaces as well. The definition of an orthogonal transformation of a subspace W ⊂ Rn is based on the
property (3.3): it is a linear transformation W → W that preserves dot products between all pairs of vectors
in W . We use (A.3) – rather than a matrix formula – to define a reflection across a hyperplane in a subspace.
3Geometrically, for n − 1 ≥ 2 if s is a reflection on H fixing the orthogonal complement of a line L in H,
then this extension of s to Rn is the reflection on Rn fixing the orthogonal complement of L in Rn .
18 KEITH CONRAD

composition of at most 2a + b = n reflections, with the understanding when n = 1 that the

identity is a composition of 0 reflections.

Example A.6. For 0 ≤ m ≤ n, we will show the orthogonal matrix

 
−1 0 ··· 0 0
 0
 −1 · · · 0 0 
 .. .. . . .. .. 
 .
 . . . . 
 0 0 ··· 1 0 
0 0 ··· 0 1

with m −1’s and n − m 1’s on the diagonal is a composition of m reflections in On (R) and
not less than m reflections in On (R).
Any reflection in On (R) has a fixed hyperplane through 0 of dimension n − 1. Therefore
a composition of r reflections in On (R) fixes the intersection of r hyperplanes through the
origin, whose dimension is at least n−r (some hyperplanes may be the same). If h ∈ On (R)
is a composition of r reflections and fixes a subspace of dimension d then d ≥ n − r, so
r ≥ n − d. Hence we get a lower bound on the number of reflections in On (R) whose
composition can equal h in terms of the dimension of {v ∈ Rn : h(v) = v}. For the above
matrix, the subspace of fixed vectors is {0}m ⊕Rn−m , which has dimension n−m. Therefore
the least possible number of reflections in On (R) whose composition could equal this matrix
is n − (n − m) = m, and this bound is achieved: the m matrices with −1 in one of the first
m positions on the main diagonal and 1 elsewhere on the main diagonal are all reflections
in On (R) and their composition is the above matrix.
In particular, the isometry h(v) = −v is a composition of n and no fewer reflections in
On (R).

Corollary A.7. Every isometry of Rn is a composition of at most n + 1 reflections. An

isometry that fixes at least one point is a composition of at most n reflections.

The difference between this corollary and Cartan’s theorem is that in the corollary we
are not assuming isometries, or in particular reflections, are taken from On (R), i.e., they
need not fix 0.

Proof. Let h be an isometry of Rn . If h(0) = 0, then h belongs to On (R) (Theorem 2.4) and
Cartan’s theorem implies h is a composition of at most n reflections through hyperplanes
containing 0. If h(p) = p for some p ∈ Rn , then we can change the coordinate system
(using a translation) so that the origin is placed at p. Then the previous case shows h is a
composition of at most n reflections through hyperplanes containing p.
Suppose h has no fixed points. Then in particular, h(0) 6= 0. By Theorem A.3 there is
some reflection s across a hyperplane in Rn such that s(h(0)) = 0. Then sh ∈ On (R), so by
Cartan’s theorem sh is a composition of at most n reflections, and that implies h = s(sh)
is a composition of at most n + 1 reflections. 

The proof of Corollary A.7 shows an isometry of Rn is a composition of at most n

reflections except possibly when it has no fixed points. Then n + 1 reflections may be
required. For example, when n = 2 nonzero translations and glide reflections have no fixed
points, and the first type requires 2 reflections while the second type requires 3 reflections.
 ISOMETRIES OF Rn 19

References
[1] S. H. Friedberg, A. J. Insel, and L. E. Spence, “Linear Algebra,” 4th ed., Pearson, Upper Saddle River
NJ, 2003.
[2] L. Rudolph, “The Structure of Orthogonal Transformations,” Amer. Math. Monthly 98 (1991), 349–352.
[3] A. Vogt, “On the linearity of form isometries,” SIAM Journal on Applied Mathematics 22 (1972), 553–
560.