Grassmannian as a metric space

Subhajit Mishra
Department of Physics, IISER Kolkata
Under the supervision of
Dr. Somnath Basu
Department of Mathematics and Statistics, IISER Kolkata

November 26, 2016

1 Introduction
Definition 1.1. (Metric Space) Let X be a set. A function d : X × X → R is called metric if the
following holds:

(i) d(x, y) ≥ 0 ∀x, y ∈ X and d(x, y) = 0 ⇐⇒ x = y;

(ii) d(x, y) = d(y, x) (symmetric);

(iii) d(x, y) ≤ d(x, z) + d(z, y) ∀x, y, z ∈ X (triangle inequality).

The set X equipped with a metric is called metric space, denoted by (X, d).

Definition 1.2. (Grassmannian of k-planes) Grassmannian of k-planes is a set of all k dimensional

subspaces of Rn . We denote it by Gk (Rn ). From now on I will call any such as Grassmannian.

We define d : Gk (Rn ) × Gk (Rn ) → Rn such that

d(L, L0 ) = sup d(x, L0 ) (1)

x∈L∩S n−1

where d(x, L0 ) = inf y∈L0 dE (x, y), with dE (x, y) as the usual euclidean metric in Rn .
We will show that (Gk (Rn ), d) is a metric space and the induced topology has properties -
compactness and path-connectedness.

2 Proof of (Gk (Rn ), d) as a metric space

In this section we will show that Gk (Rn ) is a metric space equipped the function d defined in Eqn
(1). First, we show that d is a metric.

Claim. d is a metric

Proof. (i) We begin by showing that the first property of metric holds.
First we prove a result, which will be used later.

Result 2.1. Let V and W be two k dimensional subspaces of a vector space and V ⊆ W .
Then V = W .

1
 Proof. Let dim V = dim W = k and Bv = v1 , v2 , ..., vk be a basis of V . As V ⊆ W , Bv is a
list in W . Moreover, it is a linearly independent list in W (since, it is a basis of V ).
Now, we can make Bv a basis of W by adding some vectors of W in Bv . But as dim W = k,
any basis of W must haveP length k. So, we can conclude Bv is also a basis of W . This
implies, ∀w ∈ W , w = ki=1 ai vi ∈ V (as Bv is basis of V ). So, W ⊆ V . Hence, V = W .

Now,
dE (x, y) ≥ 0; (as it is the usual metric in Rn )
=⇒ inf 0 dE (x, y) ≥ 0
y∈L

=⇒ sup inf 0 dE (x, y) ≥ 0

x∈L∩S n−1 y∈L
0
=⇒ d(L, L ) ≥ 0
Since L and L0 are are arbitrary, it is true for all L and L0 in Gk (Rn ).
Let d(L, L0 ) = 0. This implies:
sup d(x, L0 ) = 0
x∈L∩S n−1
=⇒ d(x, L0 ) = 0, ∀x ∈ L ∩ S n−1
=⇒ inf 0 dE (x, y) = 0
y∈L

=⇒ dE (x, y0 ) = 0, (for some y0 ∈ L0 , by Claim 2.4, see next page)

=⇒ x = y0
=⇒ x ∈ L0
=⇒ L ⊆ L0 .

By Result 2.1, we get L = L0 . So, d(L, L0 ) = 0 =⇒ L = L0 .

Now, consider d(L, L) = supx∈L∩S n−1 inf y∈L dE (x, y). Fix x0 ∈ L ∩ S n−1 . So,
dE (x0 , x0 ) = 0;
=⇒ inf dE (x0 , y) = 0; (as dE (x, y) ≥ 0)
y∈L

=⇒ sup inf dE (x0 , y) = 0;

x0 ∈L∩S n−1 y∈L
=⇒ d(L, L) = 0.

So, d(L, L0 ) ≥ 0 and d(L, L0 ) = 0 ⇐⇒ L = L0 .

(ii) To show that the function d is symmetric, first we show that the supremum is actually
achieved i.e., supx∈L∩S n−1 d(x, L0 ) = d(x0 , L0 ), for some x0 ∈ L ∩ S n−1 .
Lemma 2.2. The function d : Rn → R such that d(x, L0 ) = inf y∈L0 dE (x, y) is continuous in
Rn .

Proof. Let {xn }n∈N be a sequence in Rn , which converges to x0 ∈ Rn i.e., ∀ > 0, ∃N ∈ N

such that ∀n ≥ N , dE (xn , x0 ) < .
We fix n. d(xn , L0 ) = inf y∈L0 dE (xn , y) implies: given  > 0, ∃y ∈ L0 such that d(xn , L0 ) ≤
dE (xn , y) ≤ d(xn , L0 ) + . Now,
d(x0 , L0 ) ≤ dE (x0 , y) ≤ dE (x0 , xn ) + dE (xn , y), (triangle inequality)
≤ dE (x0 , xn ) + d(xn , L0 ) + 
< d(xn , L0 ) + 2, (∀n ≥ N )
=⇒ d(x0 , L0 ) − d(xn , L0 ) < 2, (∀n ≥ N and ∀ > 0)

2
By interchanging x0 and xn we will get d(xn , L0 ) − d(x0 , L0 ) < 2, ∀n ≥ N and ∀ > 0. So,
|d(xn , L0 ) − d(x0 , L0 )| < 2. Hence d(x, L0 ) is continuous at x0 . Since it is continuous for any
x0 in Rn , it is continuous in Rn .

Proposition 2.3. Let g : X → R be a continuous function and X be compact. Then

supx∈X g(x) ∈ g(X).

Proof. Since g is continuous and X is compact, g(X) ⊆ R is compact. So, g(X) is closed in
bounded (Heine-Borel Theorem). As g(X) is bounded, sup g exists, and as g(X) is closed,
sup g ∈ g(X).

Now, L ∩ S n−1 = S k−1 is compact. We restrict the function d(x, L0 ) to L ∩ S n−1 . So, d(x, L0 )
is continuous in L ∩ S n−1 (subset of Rn ). By Proposition 2.3, supremum is achieved. So,
d(L, L0 ) = supx∈L∩S n−1 d(x, L0 ) = d(x0 , L0 ), for some x0 ∈ L ∩ S n−1 .
Now, d(x0 , L0 ) = inf y∈L0 dE (x0 , y).

Claim 2.4. inf y∈L0 d(x0 , y) = dE (x0 , y0 ), for some y0 ∈ L0 , and y0 is unique.

Proof. Let y0 be the orthogonal projection of x0 into L0 . So, hx0 − y0 , y0 i = 0. Moreover, as

x0 − y0 is perpendicular to the plane L0 , we get hx − y0 , vi = 0, ∀v ∈ L0 .
Let {vi }ki=1 be an orthonormal basis of L0 . Since y0 ∈ L0 , y0 = ki=1 ai vi , ai ∈ R. We extend
P
this list to a basis {v1 , ..., vk , uk+1 , ..., un } of Rn . By Gram-Schimdt orthogonalization, we
get {vi }ni=1 , an orthonormal basis of Rn .
Now, x0 ∈ Rn =⇒ x0 = ni=1 bi vi , where bj ∈ R.
P

We take v = vj , for j = 1, 2, ..., k.

hx − y0 , vj i = 0
* n k
+
X X
=⇒ bi vi − ai vi , vj = 0
i=1 i=1
k
* +
X
=⇒ (bi − ai )vi , vj = 0, (hvi , vj i = 0, for i 6= j)
i=1
=⇒ bj = aj (for j = 1, 2, ..., k)
Pk
So, y0 = i=1 bi vi .
Pk
Let y00 ∈ L0 and y00 6= y0 . So, y00 = i=1 ci vi , ci ∈ R. Now,
v
u k n
uX X
0
dE (x0 , y0 ) = t 2
(bi − ci ) + b2i
i=1 i=k+1
k
X
=⇒ (dE (x0 , y00 ))2 = (bi − ci )2 + (dE (x0 , y0 ))2
i=1
=⇒ dE (x0 , y00 ) > dE (x0 , y0 )

Since y00 is arbitrary, we can say that inf y∈L0 d(x0 , y) = dE (x0 , y0 ).
To show that y0 is unique, let us assume ∃ỹ ∈ L0 such that hx0 − ỹ, ỹi = 0. Now, x0 =
y0 + y0⊥ = ỹ + ỹ ⊥ . This gives, y0 − ỹ = ỹ ⊥ − y0⊥ . Since y0 − ỹ ∈ L0 and ỹ ⊥ − y0⊥ ∈ L0 ⊥ and
both are equal, y0 − ỹ = 0 =⇒ y0 = ỹ. Hence, y0 is unique.

3
 So, d(L, L0 ) = d(x0 , L0 ) = dE (x0 , yx0 ), where yx0 is the orthogonal projection of x0 into L0 .
Now,
d(L, L0 ) = dE (x0 , yx0 )
= dE (yx0 , x0 ), (symmetry)
= dE (ŷ0 , xŷ0 ) (using properties of congruent triangles)
= d(ŷ0 , L), (since, infimum is achieved)
≤ sup d(y, L) = d(L0 , L).
y∈L0 ∩S n−1

where ŷ0 is the unit vector in the direction of yx0 and xŷ0 is the orthogonal projection of ŷ0
on L.
Now, interchanging L0 and L we get d(L0 , L) ≤ d(L, L0 ). So, d(L, L0 ) = d(L0 , L); (symmetric).
(iii) Triangle inequality:
Let L, L0 , L00 ∈ Gk (Rn ). We need to show that d(L, L00 ) ≤ d(L, L0 ) + d(L0 , L00 ).

Figure 1: for kyx k =

6 0

Take x ∈ L ∩ S n−1 such that d(L, L00 ) = dE (x, yx00 ), where yx00 = PL00 (x) is the orthogonal
projection of x on L00 (see Figure 1, as supremum and infimum are achieved). Using the
Pythagoras Theorem we get,
kxk2 = kyx k2 + kx − yx k2
=⇒ kyx k2 + kx − yx k2 = 1 (since x is a unit vector)
=⇒ kyx k ≤ 1; (as kx − yx k2 ≥ 0)
1
Now, ŷx = λyx , where λ = kyx k ≥ 1 (kyx k =
6 0). So,

kŷx − PL0 (ŷx )k = |λ|kyx − PL0 (yx )k (as projection map P is linear)
≥ kyx − PL0 (yx )k

The distance between ŷx and ŷx0 = PL0 (ŷx ) is always greater than the distance between yx
and yx0 . Now, translate the vector ŷx − PL0 to yx so that the line starting from yx intersects
L00 at yx0 .
Since dE (x, yx00 ) is the shortest distance between L and L00 , we have
dE (x, yx00 ) ≤ dE (x, yx0 )
≤ dE (x, yx ) + dE (yx , yx0 ); (triangle inequality)
≤ dE (x, yx ) + dE (ŷx , ŷx0 ); (as dE (yx , yx0 ) ≤ dE (ŷx , ŷx0 ))

4
 Since we have started with the x for which the supremum is achieved, from the above relation
we have d(L, L00 ) ≤ d(L, L0 ) + d(L0 , L00 ).
For the case, where kyx k = 0 i.e., two of the planes are orthogonal to each other (say L
and L0 , see Figure 2); we have d(L, L00 ) ≤ 1 = d(L, L0 ) (since, x is a unit vector). So,

Figure 2:

d(L, L00 ) ≤ d(L, L0 ) + d(L0 , L00 ) (as d(L0 , L00 ) ≥ 0).

We have shown that the function d satisfies all the properties of metric. Hence, d is a metric
and (Gk (Rn ), d) is a metric space.

The metric, d is uniformly bounded by 1. Since, the distance between any two k-planes is given by
the distance between a unit vector and the orthogonal projection of the unit vector. The distance
is 1 when there are two k-planes perpendicular to each other.

3 Some topological properties of Grassmannian

The induced topology on Grassmannian has some nice topological properties - compactness and
path-connectedness. Hausdorff property of Grassmannian is evident (as it is a metric space). We
check the two other topological properties, mentioned here.

Proposition 3.1. Gk (Rn ) is compact.

Proof. Let f : O(n) → Gk (Rn ) be a function defined as f (A) = span(A1 , A2 , ..., Ak ), where Ai is
a coloumn vector of A. If we can show that f is onto and continuous, then we are done.
Claim 3.2. (f is onto) ∀L ∈ Gk (Rn ), ∃A ∈ O(n) such that f (A) = L.

Proof. Let BL = {vi }ki=1 be an orthonormal basis of L. Since BL is a linearly independent list, we
extend it to {v1 , v2 , ..., vk , uk+1 , ..., un }, a basis of Rn . Using Gram-Schimdt orthogonalization we
get an orthonormal basis {vi }ni=1 of Rn .
So taking A = [v1 v2 ... vn ], we have f (A) = span(v1 , v2 , .., vk ) = L. As vi ’s are orthonormal,
A ∈ O(n). Since L is arbitrary, f is onto.

Claim 3.3. (f : O(n) → Gk (Rn ) is continuous) ∀ > 0, ∃δ > 0 such that ∀A ∈ O(n), ∀B ∈ O(n)
if dE (A, B) < δ then d(f (A), f (B) < .
(Note). We have taken the metric in O(n) to be euclidean because we can identify an element of
2
O(n) as a vector of Rn .

5
Proof. Let A = [aij ] and B = [bij ] be elements of O(n). Now,
n
X
(dE (A, B))2 = (aij − bij )2 < δ 2
(i,j)=1
X n n
X
2
=⇒ (ai1 − bi1 ) + ... + (ain − bin )2 < δ 2
i=1 i=1
n
X
=⇒ (aij − bij )2 < δ 2 ; since (aij − bij )2 ≥ 0, for each j = 1, 2, ..., n
i=1
=⇒ dE (Aj , Bj ) < δ, (for each j = 1, 2, ..., n)
Since supremum and infimum both are achieved, we have d(f (A), f (B)) = supx∈f (A)∩S n−1 d(x, f (B))
= d(x, f (B)) = dE (x, Pf (B) (x)), for some x ∈ f (A) ∩ S n−1 , where P is the orthogonal projection
operator. Now, x = ki=1 λi Ai , with ni=1 λ2i = 1. So,
P P

d(f (A), f (B)) = dE (x, Pf (B) (x))

k k
!!
X X
= dE λi Ai , Pf (B) λ i Ai
i=1 i=1
k k
!
X X
= dE λ i Ai , λi Pf (B) (Ai ) , (since, P is linear)
i=1 i=1
k
X  
≤ dE λi Ai , λi Pf (B) (Ai ) , (using triangle inequality)
i=1
Xk  
= |λi |dE Ai , Pf (B) (Ai ) , (as dE is the euclidean norm)
i=1
Xk    
≤ |λi |dE (Ai , Bi ); as dE Ai , Pf (B) (Ai ) is the infimum
i=1
√ k
X √
< kδ (as dE (Ai , Bi ) < δ and using Cauchy-Schwarz inequality, |λi | ≤ k)
i=1

=  (taking δ = √ )
k
Since f is continuous for all A, B in O(n), it is continuous in O(n).

As, f : O(n) → Gk (Rn ) is onto and continuous, and O(n) is compact, we have Gk (Rn ) is compact.

Proposition 3.4. Gk (Rn ) is path-connected.

Proof. Let f : SO(n) → Gk (Rn ) is a function defined as f (A) = span(A1 , A2 , ..., An ).

As SO(n) ⊂ O(n) and f is continuous on O(n) (shown in the previous section), f is continuous
on SO(n). It is sufficient to show that f is onto.
Going by the same logic as in the proof of Claim 2, we get an orthogonal matrix
A = [v1 v2 ... vn ]. If det(A) = 1, then A ∈ SO(n) and we are done.
If det(A) = −1, we do a linear transformation, to get A0 such that
 
−1 0 0 ... 0
 0 1 0 ... 0
A0 = [v1 v2 ... vn ]  .  = [−v1 v2 ... vn ]
 
 .. ..
. 
0 0 0 ... 1

6
Now, det(A0 ) = 1. So, A0 ∈ SO(n). Hence, f is onto.

We know that continuous function preserves path-connectedness. Since SO(n) is path-connected

and f : SO(n) → Gk (Rn ) is continuous and onto, Gk (Rn ) is path-connected.