Lecture 5 - Hausdorff and Gromov-Hausdorff Distance
August 1, 2011
1 Definition and Basic Properties
Given a metric space X, the set of closed sets of X supports a metric, the Hausdorff metric.
If A is a set in X and r > 0, we define the r-thickening, or r-neighborhood, of A to be the
set A(r) defined by
[
A(r) = Bx (r) (1)
x∈A
where Bx (r) is the (open) ball of radius r about x. If A, B ⊂ X are closed sets, define their
Hausdorff distance dH (A, B) to be the number
dH (A, B) = inf { r > 0 | B ⊂ A(r) and A ⊂ B (r) }. (2)
Recall that the infimum of an empty set is regarded to be +∞. A equivalent definition is
as follows. Given a point p ∈ X and a closed set A ⊂ X, define
d(p, A) = inf dist(p, y). (3)
y∈A
Then the Hausdorff distance is
dH (A, B) = max sup d(x, B), sup d(y, A) (4)
x∈A y∈B
That is, dH (A, B) is the farthest distance any point of B is from the set A, or the farthest
any point of A is from B, whichever is greater. Again, this could be infinite.
Theorem 1.1 If (X, d) is a bounded metric space, the set of closed sets of X is itself a
metric space with the Hausdorff metric.
Pf We verify the metric space axioms. First, the symmetry of dH is clear by definition.
Second, dH satisfies the triangle inequality because if C is in the r-neighborhood of B and
1
B is in the s-neighborhood of A, then C is in the (r + s)-neighborhood of A. Likewise A is
in the (r + s)-neighborhood of C. Thus d(A, C) ≤ d(A, B) + d(B, C). Finally dH (A, B) = 0
implies A ⊆ B = B, because if B is in every r-neighborhood of A then every point of A is
a limit point of B. Likewise B ⊆ A = A.
If X is not bounded, the metric space axioms continue to hold except that possibly
closed sets A and B exist with dH (A, B) = ∞. This could be rectified by restricting to
compact subsets of X, although this is not natural in some cases.
2 Compactness Properties
Let (X, d) be a metric space and denote the set of closed subsets of X by C(X) (or just C
for short). Given a closed set A and a number r, let BA (r) be the set of all D ∈ C with
dH (B, A) < r (that is, the r-ball around A in C). Since dH is a metric on C, we know that
the balls BA (r) are open, and form a neighborhood base.
Obviously the balls with rational radius also form a base, so the topology on C induced
by dH is first countable. All metric spaces are Hausdorff, so (C, dH ) is Hausdorff. One can
state this directly: since distinct closed sets are separated by a finite distance, say , so the
balls of radius, say, /4 around each is disjoint.
Theorem 2.1 If (X, d) is a compact metric space, then (C(X), dH ) is compact.
Pf
The proof can go as in the proof of Gromov’s precompactness theorem; we leave it as
an exercise.
It should be clear that if (X, d) is non-compact, then (C(X), dH ) is non-compact. This
can be seen by the existence of an obvious isometric embedding X ,→ C(X), and by noting
that if a sequence in X converges in C(X), its limit must be a point, and therefore again an
element of X.
If (X, d) is bounded, then it is locally compact if and only if (C(X), dH ) is locally
compact. It can be proven that (C(X), dH ) is paracompact whenever (X, d) is bounded.
The proofs are left as exercises.
In sharp contrast, if (X, d) is unbounded, then (C(X), dH ) need not be locally compact
nor even locally paracompact. For instance if the base space X is unbounded and nondiscrete
(it has the property that, given any point x ∈ X and any number > 0, there is a point
y ∈ X with d(x, y) < ), then it is not locally compact. As an example, we will will
show that R is not locally compact. Let A = [0, ∞) be the half-line, and consider its r-
neighborhood BA (r) (wlog assume r < 21 ). Define the Ai inductively by setting A0 = A and
Ai = Ai−1 \ (i, i + r/2). We have dH (Ai , Aj ) = r/2 for any i 6= j, so there are no Cauchy
subsequences and therefore no convergent subsequences.
2
A topology does exist on C(X) that is both locally compact and compact, whether X
is bounded or not. Let a base for this topology be set of the form NK, (A), where K ⊂ X
is compact, A ⊂ X is closed, and > 0, where we define
NK, (A) = { B ∈ C(X) | dH (A ∩ K, B ∩ K) < }.
This topology on C(X) is called the pointed Hausdorff topology. If X is compact, it is the
metric topology. If X is noncompact, this topology is not induced by any metric.
3 The Gromov-Hausdorff distance
The Gromov-Hausdorff distance significantly extends the idea of the Hausdorff distance.
Given two closed metric spaces A and B, we define
dGH (A, B) = inf dH (fA→X (A), gB→X (B)) (5)
f,g
where the notation fA→X (resp. gB→X ) denotes an isometric embedding of A into some
metric space X (resp. an isometric embedding of B into X) and the infimum is taken
over all possible such embeddings. Note that dGH could well be infinty; however it is
clearly symmetric. To show that dGH (A, B) = 0 iff A and B are isometric, we first give an
equivalent definition of dGH .
Proposition 3.1 The Gromov-Hausdorff distance dGH (X, Y ) is the
` infimum of the Haus-
dorff distances between X and Y taken among all metrics on X Y that restrict to the
given metrics on X and on Y .
Pf Exercise.
Proposition 3.2 If (X, dX ) and (Y, dY ) are metric spaces that admit compact exhaustions,
and that dGH (X, Y ) = 0. Then (X, dX ) and (Y, dY ) are isometric.
Pf
If X and Y are isometric then clearly dGH (X, Y ) = 0.
Conversely assume dGH (X, Y ) = 0, and for the moment
` assume X and Y are compact.
Then there is a sequence of distance functions di on X Y with di |X = dX and di |Y = dY
so that di,H (X, Y ) → 0. Let j > 0 be a sequence that converges to 0. For each j construct
finite sets of points Xj = {xk } and Yj = {yk } with the following properties: Xj is j -dense
in X, Yj is j -dense in Y , and for large enough i the sets Xj and Yj are j -close S in the
Hausdorff metric. WeS also require that Xj ⊂ Xj+1 and Yj ⊂ Yj+1 , so that X = j Xj is
dense in X and Y = j Yj is dense in Y .
3
Now consider the distance functions {di } restricted to Xj ∪Xj . Because Xj ∪Xj is finite,
a subsequence dij converges to a limiting pseudometric dj . Passing to refined subsequences
as j increases and taking a diagonal
` subsequence, we get convergence`to a pseudometric d
on X ∪ Y, a dense subset of X Y , and therefore convergence on X Y .
Given any j , a given point x ∈ X is j -close to a point xj ∈ X , which is j -close to
a point of yj ∈ Y. Taking a limit y = limj yj we have that d(x, y) = 0; because X is a
metric space, this point x is unique (similarly given a point y ∈ Y we can find a unique
point x ∈ X with d(x, y) = 0). Finally send x to the unique point y ∈ Y with d(x, y) = 0.
The fact that this is an isometry follows from the triangle inequality: if x1 , x2 ∈ X are
sent to y1 , y2 ∈ Y , respectively, then
dX (x1 , x2 ) = d(x1 , x2 ) ≤ d(x1 , y1 ) + d(y1 , y2 ) + d(y2 , x2 ) = dY (y1 , y2 )
dY (y1 , y2 ) = d(y1 , y2 ) ≤ d(y1 , x1 ) + d(x1 , x2 ) + d(x2 , y2 ) = dX (x1 , x2 ),
so that dX (x1 , x2 ) = dY (y1 , y2 ).
4 Gromov-Hausdorff Approximations
We mention what is often a more useful formulation of the Gromov-Hausdorff distance.
A map f : X → Y (not necessarily continuous) between metric spaces is called an -
GHA (for “Gromov-Hausdorff approximation”) if |dY (f (x1 ), f (x2 )) − dX (x1 , x2 )| < for
all x1 , x2 ∈ X, and Y is in the -neighborhood of f (X). We can define a new distance
function between metric spaces, called dd
GH , by setting
0
GH (X, Y ) = inf{ > 0 | there are −GHA s f : X → Y and g : Y → X }.
dd
It is a simple exercise to prove that this is a metric: if there is an 1 -GHA f : X → Y and
an 2 -GHA g : Y → Z, then the composition satisfies
|dZ (gf (x1 ), gf (x2 )) − dX (x1 , x2 )|
≤ |dZ (gf (x1 ), gf (x2 )) − dY (f (x1 ), f (x2 ))| + |dY (f (x1 ), f (x2 )) − dX (x1 , x2 )|
≤ 1 + 2
and it is also easy to show that the (1 + 2 )-neighborhood of f g(X) is Z. Taking infima,
we have that dd GH (X, Z) ≤ dGH (X, Y ) + dGH (Y, Z).
d d
Proposition 4.1 The metrics dd
GH and dGH are equivalent (though they are not the same).
Proof. Exercise.
4
5 Compactness Properties
Proposition 5.1 The Gromov-Hausdorff topology on the set of compact metric spaces is
second countable.
Pf Exercise. (Hint: If a topology is Hausdorff and separable it is second countable.)
Lemma 5.2 (Gromov’s Precompactness Lemma) Let N : N → N be monotonic. As-
sume M is a collection of metric spaces so that each M ∈ M has a 1j -dense discrete subset
of cardinality ≤ N (j). Then M is precompact.
Proof. Let {Mi } ⊂ M, and let M̃i,j ⊂ Mi be a 1j -dense subset of cardinality ≤ N (j). By
Pj
replacing N (j) with i=1 N (i) we can assume that M̃i,j ⊂ M̃i,j+l . Fixing j and letting
i → ∞ we get convergence of M̃i,j along a subsequence to a space M̃j . Passing to further
refinements of the subsequence and taking a diagonal sequence, we get a sequence of distance
S
functions dk that converge on each M̃j , and therefore on M̃ = j M̃j . Now given > 0
there is an i so that M̃i is -close to M̃ , and there is a j so that Mi,j is -close to both M̃i
and to Mi . Thus Mi converges to M̃ .
In general the topology associated to the Gromov-Hausdorff distance is neither locally
compact nor locally paracompact. To redress this we define the pointed Gromov-Hausdorff
topology, which is locally compact and compact. On the space of compact metric spaces, this
will be the same as the original Gromov-Hausdorff topology. However the pointed topology
is not induced by any norm.
The pointed Gromov-Hausdorff topology is defined on the set of pointed metric spaces
(defined to be pairs (A, p, d) where (A, d) is a closed metric space and p ∈ A). A local base
for this topology are the sets of the form NK, (A) (where A is closed, K ⊂ A is compact
and p ∈ K, and > 0); we define NK, (A) to be the set of pointed closed sets (B, q) so that
there exists a compact subset J ⊂ B, q ∈ J, and so that there are isometric embeddings
f : A ∩ K → X and g : B ∩ J → X into some space X so that f (p) = g(q) and the Hausdorff
distance satisfies dH (f (A ∩ K), g(B ∩ J)) < .
6 Exercises
1) Give an example of a bounded metric space that is not locally compact.
2) Prove that if (X, d) is bounded, then it is locally compact iff (C(X), dH ) is locally
compact.
3) Prove that if (X, d) is bounded, then (C(X), dH ) is paracompact.
5
4) Let (R, d) be the real line with the standard metric. Construct an uncountable discrete
subset of (C(R), dH ).
5) Prove that if (R, d) is the real line with the standard metric, the topology induced by
(C(R), dH ) is not locally paracompact (note that dH is not a metric on C(R)).
6) Prove Theorem 2.1, namely that a compact metric space induces a compact Hausdorff
metric.
7) Prove Theorem 4.1, namely that dd
GH is equivalent to dGH .
8) Prove that the pointed Gromov-Hausdorff topology is not second countable.
9) In a previous lecture, we constructed a sequence of metrics gδ on a nilmanifold Γ\N ,
where N was the Heisenberg group and Γ was the integer lattice. Prove that given
any , there is a δ to that the map Γ\N 7→ pt is an -GHA.
10) In a previous lecture, we constructed a family of metrics gt , the Berger metrics, on
the manifold S3 . Prove that the Hopf map S3 7→ S2 induces a πt-GHA from (S3 , gt )
to (S2 , 4gS2 ) where gS2 is the standard metric on S2 .
