How big is that cookie?

The Integral Geometric approach to

geometrical quantities
Abstract. Integral geometry studies the link between expectation
of random variables and geometrical quantities like length, area or
curvature. This thesis focuses on how expectation of random vari-
ables can be used to define reasonable notions of geometrical size.
The simple idea of looking at the shadow when a compact convex
body in Rn is orthogonally projected onto a random k-dimensional
subspace will generate a collection of “continuous invariant valua-
tions”. A theorem of Hadwiger concludes that this collection spans
the vector space of all continuous invariant valuations, and thus
integral geometry elegantly provides a complete description about
the size of a compact convex body. Furthermore, some of these
valuations turn out to be extensions of familiar geometrical con-
cepts like surface area, but escape the typically required continuity
or smoothness conditions and offer an alternative interpretation of
these concepts.
In choosing a k-dimensional subspace at random, one needs to
specify a probability measure that is “invariant under the symme-
tries of Euclidean space” so that the size of an object is unchanged
after it is displaced by a Euclidean motion. We explore the pos-
sibility of imposing measures without such restrictions to obtain
new kinds of geometries, and perform this exploration on a generic
smooth manifold. Finally, we look at how a variant of the integral
geometric ideas explored permits the study of geometrical quanti-
ties of a different flavor: various forms of curvature.
Harvard University 2015
Barry Jia Hao Tng

Advisor:
Oliver Knill
 1

Acknowledgements
First, I would like to thank my family, especially my younger brother, for always
being supportive of me and cheering me on during difficult times. To my parents,
I am proud to say that I have learnt a tremendous amount through my years in
school, and for that I have to thank you for your love and the environment you
created for me.
Next, I would like to thank my advisor, Oliver Knill, for introducing me to this
topic that I had no prior exposure to but turned out to be an elegant and fascinating
area of study. I thank him for all his patience in advising me and infecting me with
all the enthusiasm he has for mathematics.
I have had the greatest fortune of being taught through the years by some of
the most dedicated and amazing teachers, be it from Harvard or Raffles. I am
delighted to have learnt a great deal of mathematics from Sarah Koch, Dennis
Gaitsgory, Curtis McMullen, Jacob Lurie, Alex Perry, Junecue Suh, Melody Chan,
Sukhada Fadnavis, Siu Cheong Lau, Mark Kisin, and Shlomo Sternberg. I would
also like to thank Joe Blitzstein and David Morin for some of the most amazing
courses I have taken at Harvard.
Finally, I would like to thank my most precious friends who always stand by me
- Zhang Wei, An, Chris, Diana and Dan. Each of you have a unique sense of humor
that never fails to cheer me up, and all of you have inspired me with your outlooks
about life. I would also like to thank Bang for helping me with the figures, and
numerous friends who offered to look over this thesis, especially Jeremy.
2

Contents
Acknowledgements 1
1. Introduction 3
1.1. Buffon and the Lens of Probability 3
1.2. Bertrand and Caution with Probability 5
1.3. Integral Geometry 8
1.4. Outline of Thesis 9
2. Invariant Measures 10
2.1. Theory of Lie Groups and Homogeneous Spaces 10
2.2. Grassmannians 13
2.3. k-planes 14
2.4. Example - Buffon’s Needle Problem in Rn 15
3. Notions of Size 18
3.1. “Reasonable” notion of size 18
3.2. Intrinsic Volumes 20
3.3. Hadwiger’s Theorem 22
3.4. Steiner’s formula 23
4. Lengths on Manifolds 29
4.1. From Needles to Noodles 30
4.2. Extension to Manifolds 30
5. Total Curvature Measures 34
5.1. Weyl’s Tube Formula 34
5.2. Local Steiner’s Formula 35
5.3. V -measures and Integral Geometry 37
6. Recent Developments 41
References 43
1. Introduction
There is no royal road to geometry.
– Euclid of Alexandria

In the 3rd Century BC, Euclid wrote one of the most influential textbooks of
all time – The Elements. In the treatise, Euclid revolutionized geometry by in-
troducing an axiomatic approach. In the two millennia that have since passed,
geometrical concepts like length, area and volume continue to be of relevance to
the human mind but have been subjected to much higher levels of abstraction and
complexity in order to meet the standards of modern mathematics. For example,
the definition of surface area can only be made for certain subsets of R3 , such as a
smooth surface, and may involve technical machineries such as parametrization and
working with integration of forms. Other geometrical concepts such as curvature
are only given rigorous definitions on a relatively recent timescale using the ab-
stract setting of manifold theory and differential geometry, and also involve highly
technical formulations.
Following the wise words of Euclid, one might gain new insights by exploring an
alternative road to understanding some of these geometrical quantities, especially
in light of the high level of abstractions and technicalities involved in the definitions.

1.1. Buffon and the Lens of Probability. In 1733, the French mathematician
Comte de Buffon first made the connection between probability theory and geom-
etry with his famous Needle Problem (problem 1.1.1), opening the possibility of
using a different approach to understanding geometry – using probability. Here is a
quick recap of the problem and his solution published 4 years later, which will also
make concrete what is meant by a possible probabilistic approach to understanding
geometry.
Problem 1.1.1. (Buffon’s Needle Problem) Let the xy-plane be ruled with parallel
lines of 1 unit apart – for concreteness, take the system of lines of the form y = m
where m ∈ Z. Drop a needle (i.e. line segment with endpoints) of length l randomly
onto the plane. What is the probability that the needle will cross at least one of the
lines? For the purpose of this discussion, we shall only consider the case 0 < l < 1,
so that the needle can only cross at most one of the lines.
In formulating the problem, one needs to be mathematically precise about what
is meant by dropping the needle randomly onto the plane. To do this, note that it
is sufficient to describe the position of the needle by a, the vertical distance between
the centre of needle to the closest line, and θ, the acute angle formed by the needle
and the vertical direction. We shall reasonably choose these random variables to
independently follow a uniform distribution, specifically:
a ∼ Unif [0, 1/2] and θ ∼ Unif [0, π/2]
4

1
l
aθ cos θ
2

a l
θ cos θ
1 2

Figure 1.1.

This means that the joint probability density function of (a, θ) is

(
4/π 0 ≤ a ≤ 1/2 and 0 ≤ θ ≤ π/2
0 otherwise

Solution. A simple geometric analysis (figure 1.1) shows that the needle crosses a
line if and only if
l
a ≤ cos θ
2
The probability that our random variables satisfy the above inequality is given by
Z π2 Z 2l cos θ
4 2l
da dθ =
0 0 π π
This solves the Buffon’s Needle Problem (at least for 0 < l < 1). 

There is an alternative interpretation of the above result. Let X be the random

variable denoting the number of intersections between the needle and the system
of lines. In our case of 0 < l < 1, X can only take two possible values, 0 or 1. This
means that E[X] is the probability that the needle intersects some line. Therefore1,
2l
(1.1) E[X] =
π
It is easy to show that equation (1.1) holds for all l > 0. Namely, mentally
consider the needle to be composed to shorter needles of length l1 , . . . , ln pieced
together2, where 0 < li < 1 and l1 + · · · + ln = l. Define Xi to be the number
of intersections between the ith piece and the system of lines, which means X =

1Note that this result holds even if our needle does not have one or both its endpoints included.
2The careful reader would worry about what happens at the gluing points between two short
needle pieces, since they appear to be counted twice. However, by the previous footnote, we can
ensure that each point of a long needle is only counted once by disregarding, where necessary, one
or more endpoints of our short needle pieces.
 5

X1 + · · · + Xn . Now, linearity of expectations holds regardless of the dependency

between the Xi ’s, so3
" n # n n
X X X 2li 2l
E [X] = E Xi = E [Xi ] = =
i=1 i=1 i=1
π π
This is a remarkable result. It states that the length of a needle can be computed
(up to a scaling factor of π/2) by looking at the average number of intersections with
the system of lines. This is a glimpse of how the expectation of a random variable
can be used to calculate a geometrical quantity. We shall need this relation later,
so let us record our discussion down as a proposition.
Proposition 1.1.2. Under the set-up of problem 1.1.1 (and the accompanying
definition of “random”), the expected number of intersections between the needle
and the system of lines is given by 2l/π.
1.2. Bertrand and Caution with Probability. A key decision made in formu-
lating the Buffon’s Needle Problem was how to place the needle randomly. The
infamous Bertrand’s paradox tells us that this is not a decision that can be ig-
nored. Although the paradox will not be discussed in detail here (see [Tis84]), it
suffices to say that Bertrand exhibited the possibility of obtaining different answers
to probability questions if the phrase “at random” can be interpreted in different
ways. In our context, if we had chosen a different definition of placing our needle
randomly, then E [X] would turn out to be different, and may in fact be unrelated
(or related in a complicated way) to l. We were fortunate that the most intuitive
choice made above turns out to give not just a relation between E [X] and l, but
in fact the relation is linear (i.e. simple!). How might one choose the probability
measure to ensure that there is a relation between E [X] and l?
To get a sense of what additional properties our probability measure should
have, consider the Buffon’s Needle Problem generalized to 3 dimensions. Here,
there are a system of parallel planes spaced 1 unit apart and a needle of length
l. Previously, as a matter of respecting the convention of the original problem, we
fixed the system of lines and placed the needle randomly. From this point on, it
will be more useful to think that there is a needle fixed in space whose length we
are interested in measuring, and we are doing this by randomly imposing a system
of parallel planes and computing the expected number of intersections, which we
will denote by E [X]. In other words, we get to choose the probability measure on
systems of planes and want a relation between E [X] and l that is true no matter
how the needle has been fixed in space.
Suppose temporarily that we always fix our needle to be parallel to the z-axis
of our xyz-space, with its middle at the origin. Here is a perfectly legitimate
probability measure to impose on our system of planes that will turn out to give
a linear relation between E [X] and l in this temporary set-up, but will later be
shown to actually be lacking. A system of planes is determined once we picked
(θ, φ, a), where the polar angle θ and azimuth angle φ determines a direction in
space for which the planes will be perpendicular to, and a is the closest distance of

3Before we can apply the result E [X ] = 2l /π, we have to show that the position of the ith
i i
piece has been placed in accordance to the (a, θ) distribution described above. Technically we
only know this for the entire needle, but it is straightforward to see that this implies the same for
the ith piece.
6

θ
a
y
φ

Figure 1.2.

the origin to a plane in the system (see figure 1.2). We set

θ ∼ Unif [0, π] , φ ∼ Unif [0, 2π] , a ∼ Unif [0, 1/2]
with these three random variables being independent.
If 0 < l < 1, then a similar geometric analysis tells us that the needle crosses
a plane if and only if the needle, when projected onto the line normal to all the
planes, crosses a plane. This happens if and only if
l
a≤ |cos θ|
2
See figure 1.3. The probability that our random variables satisfy the above inequal-
ity is given by
Z 2π Z π Z 2l |cos θ|
1 1 2l
2· · da dθ dφ =
0 0 0 π 2π π
By a similar reasoning to those leading up to theorem 1.1.2, we see that for all
l > 0, we have E [X] = 2l/π. There is indeed a linear relation between E [X] and l.
However, in general, the needle is simply fixed in space, pointing in an arbitrary
direction at an arbitrary position. We want to use a single probability measure on
our system of planes to compute the length of any fixed needle. Unfortunately, the
formula E [X] = 2l/π is no longer always true. For example, if the fixed needle
 7



l
2

 a
l
|cos θ| y
2

Figure 1.3.

is pointing along the x-direction with its middle at the origin, then by the same
projection argument a needle of length l < 1 crosses a plane if and only if
l
a ≤ sin θ|cos φ|
2
The probability that our random variables satisfy the above inequality is given by
Z 2π Z π Z 2l sin θ|cos φ|
1 1 4l
2· · da dθ dφ = 2
0 0 0 π 2π π
This time, we have E [X] = 4l/π 2 6= 2l/π.
In summary, the probability measure that was imposed failed to give any relation
between E [X] and l, beause E [X] also depends on other parameters of the needle
(e.g. its direction). There is actually a linear relation if we restrict the needles to
always be lying along a certain direction in space, but this is a silly restriction. If
only the probability measure is somehow compatible with rotation, then a single
relation holding for needles along one direction will also hold for all needles in
general. This motivates the following definition.
Definition 1.2.1. Suppose T is a set equipped with a group action by G. For a
subset S ⊆ T and an element g ∈ G, we write gS := {gt | t ∈ S}. The measure µ
of a measure space (T, Σ, µ) is said to be G-invariant (or simply invariant) if for
all S ∈ Σ, we have gS ∈ Σ and µ (S) = µ (gS).
8

The set T containing all possible systems of planes is equipped with the action
by the Euclidean group E(3). The preceding discussion motivates why any measure
we consider imposing on T should preferably be E(3)-invariant (and Σ should be
fine enough so that X : T → R ∪ {±∞} is a random variable4, i.e. measurable5).
Specifically, if for a particular fixed needle we find E [X] to be of a certain value,
then we can subject the needle to any Euclidean motion and its length is not
going to change under Euclidean motion, so E [X] had better be unchanged if we
want a relation between E [X] and l. Choosing a E(3)-invariant measure on T
will guarantee this. Of course, the relation may not be linear, but it is a relation
nonetheless. We will finish up this discussion of the Buffon’s Needle Problem once
we have more tools (example 2.4.1).

1.3. Integral Geometry. In the late 1930s, Wilhelm Blaschke published a series
of papers under the project “integral geometry”. His idea was to investigate, as
we have tried above, whether expectations of random variables could be used for
calculating and understanding geometrical quantities like length, area or curvature.
Morgan Crofton, in the late 1800s, actually preceded this effort with the discovery of
many simple relations of this flavor. However, his work and even simply the idea of
using probability to study geometrical quantities, which went by the umbrella title
“geometric probability”, came under great threat of being completely discredited
after the paradox of Bertrand attacked Crofton’s loose treatment of randomness.
With no unified or consistent approach to value one definition of randomness over
another, the theory appears highly arbitrary as to when a definition will lead to a
relation. It was Poincaré who, in 1896, kept the idea alive by suggesting that the
only definitions of randomness worth considering be limited to measures that are
invariant under any symmetry group which the interested geometrical quantity is
known to be invariant under. He coined the term “kinematic density” for what is
more commonly called an invariant measure today.
Having been revived by the rebranding of Blaschke, the field has made notewor-
thy progress in the past half a century. The most important is probably establishing
the use of homogeneous space theory as a unified approach to obtain the measure
of interest, a work done by two of Blaschke’s students, André Weil and Shiing-Shen
Chern. Integral geometry offers itself as an interesting alternative tool to study ge-
ometrical problems (for example, Milnor [Mil50, 3.1] used this to study curvature of
knots), but is also used today in stochastic geometry [SW08], computational model-
ing (for example, tomography) [KW03] and even made an appearance in statistical
physics [Mec98].
The name “integral geometry” deserves a brief comment. The goal of the field is
to link geometry with expectation of random variables, which is actually an integral
of the random variable as a measurable function over the probability space, hence
the name. In some cases, it is neater to not limit ourselves to probability spaces.
Throughout, we will abuse terminology and also use the term “expectation” of a
measurable function to denote its integral over the measure space.

4We considered the extended real numbers because the needle could very well lie in one of the
planes, in which case the number of intersections is ∞. We might be tempted to disregard this
situation by saying that it occurs with zero measure, but recall that we have yet to define our
measure on T , so the phrase ”zero measure” makes no sense.
5
The Borel σ-algebra of the extended real numbers consists of all sets S for which S ∩ R is a
Borel set in the usual sense.
 9

1.4. Outline of Thesis. Integral geometry is a rich field with its tentacles span-
ning into many areas of mathematics. This thesis will choose to focus on how
expectations can be used to define reasonable notions of geometric sizes. For ex-
ample, the “surface area” of a solid in R3 might be considered a reasonable notion
of size. However, we remarked earlier that a formal definition by conventional
methods only assigns surface area to certain subsets of space and involves certain
technical machineries to set-up fully. Using the idea of expectations, we will – in
a single breath – cleanly define an entire collection of notions of geometric size,
one of which will be reminiscence of surface area (we will, for example, realize that
there is agreement for convex polyhedrons), defined even for some subsets of R3 for
which the conventional approach does not ascribe a surface area. A theorem due
to Hadwiger will tell us why the single breath we took was enough to understand
all reasonable notions of geometric sizes. This discussion will be done in chapter 3
and is the center point of the thesis.
Chapter 2 carries out the preparatory work of establishing certain invariant
measures that we will need. In particular, for 0 ≤ k ≤ n, the Grassmannian
Gr(n, k) := {k dimensional subspaces of Rn } is acted upon by the orthogonal
group O(n) and the set of all k-planes Aff(n, k) is acted upon by the Euclidean
group E(n); we will need invariant measures for both. To do this, we will use the
theory of homogeneous spaces and treat O(n) and E(n) as Lie groups.
Chapters 4 and 5 are bonus sections where we extend particular ideas presented
in chapter 3. Treatment in these two chapters is designed to give a flavor of the
power of integral geometry beyond this thesis.
In chapter 4, we present some original ideas and pursue the thread of using
integral geometry to define the notion of length, but extended to curves on arbitrary
smooth manifolds. The chapter will allow us to better appreciate how integral
geometry can be used as the starting point for prescribing geometries, assigning
distance functions to a smooth manifold with no presupposed geometry.
In chapter 5, we discuss how the integral geometric framework can also encom-
pass the different geometrical notion of curvature.
2. Invariant Measures
In these days the angel of topology and the devil of abstract algebra fight
for the soul of every individual discipline of mathematics.
– Hermann Weyl

As we have seen, one of the key prerequisites to integral geometry is the study
of invariant measures. In this chapter, we recap some of the main takeaways from
the theory of homogeneous spaces that will be useful for integral geometry. In
particular, we will establish the existence of invariant measures for a wide class of
contexts that are frequently useful for practical applications in integral geometry.
We will also pay particular attention to the examples of Grassmannians and the
set of k-planes in Rn .
For this purpose, we will be needing some standard material from smooth man-
ifold theory, which we will be prepared to cite as long as it is discussed in [Lee12],
a standard textbook on the subject.
2.1. Theory of Lie Groups and Homogeneous Spaces. In definition 1.2.1, we
are under the most general context where there is simply a set T equipped with
a group action by G. In most applications, G can have an additional structure of
a Lie group, which will turn out to be a huge resource. Recall that a Lie group
G is a group whose elements are also points of a smooth manifold, such that the
group operation (g, h) 7→ gh and inversion map g 7→ g −1 are both smooth. There
are two examples of Lie groups that will be relevant to our purpose. They are
the orthogonal group O(n), which is the set of n × n real matrices A satisfying
AT A = AAT = I, and the Euclidean group E(n), which is the symmetry group of
n-dimensional Euclidean space (i.e. in particular including the translation group
T (n) and O(n) as subgroups). For a proof that these are Lie groups, see [Lee12,
7.27] and [Lee12, 7.32].
Let us first study the situation of a Lie group G acting on by itself via left
multiplication and see if we can obtain an invariant Borel measure on G (in the
sense of definition 1.2.1). Note that here we chose the σ-algebra of the measure
space to be Σ = B(G), where the notation B(X) will be used to denote the Borel
σ-algebra of X whenever X is a topological space.
Indeed, we have the following useful starting point from smooth manifold theory.
Theorem 2.1.1. Every Lie group, which can always be endowed with a left-
invariant orientation, has a nowhere-vanishing positively-oriented left-invariant (con-
tinuous) n-form (where n = dim G). Moreover, all left-invariant n-forms are equal,
up to constant multiples.
Proof. See [Lee12, 16.10]. 
Theorem 2.1.2. Let G be a Lie group acting on itself by left multiplication. Then
there is a measure µ on (G, B(G)) such that µ is invariant.
 11

Before proceeding with the proof, we need a few pieces of notation. We let Cc (G)
denote the space of continuous functions on G with compact support. If U is an
open set in G and f ∈ Cc (G), we use the notation
f ≺U
to mean 0 ≤ f ≤ 1 and supp(f ) ⊆ U . Also observe that if f ∈ Cc (G) and ω is
aR (continuous) n-form, then f ω is a compactly supported (continuous) n-form, so
G
f ω makes sense (after we chose an orientation for G).
Proof. Endow G with a left-invariant orientation and let ω denote a nowhere-
vanishing positively-oriented left-invariant (continuous) n-form.
Step 1: Let us first construct a candidate Borel measure. We define, for U ⊆ G
open, Z 
µ0 (U ) := sup f ω | f ∈ Cc (G), f ≺ U
G
and for an arbitrary E ⊆ G, we define
µ∗ (E) := inf {µ0 (U ) | U ⊇ E, U open}
Observe that µ0 (U ) ≤ µ0 (V ) if U ⊆ V and so µ∗ (U ) = µ0 (U ) for U open. Let us
show that µ∗ is an outer measure. For that, we need this lemma.
S∞
Lemma: If U1 , U2 , . . . is a sequence of open sets and U := j=1 Uj , then
P∞
µ0 (U ) ≤ j=1 µ0 (Uj ).
Proof of Lemma: Let f ∈ Cc (G) be such that f ≺ U . By compactness of
SN SN
supp(f ), we can pick N such that supp(f ) ⊆ j=1 Uj . Since j=1 Uj is a smooth
manifold, by the Existence of Partitions of Unity theorem [Lee12, 2.23], we can find
SN PN
smooth functions g1 , . . . , gN : j=1 Uj → R such that gj ≺ Uj and j=1 gj ≡ 1 on
SN PN
j=1 Uj . Then f = j=1 f gj and f gj ≺ Uj , so
Z Xn Z n
X ∞
X
fω = f gj ω ≤ µ0 (Uj ) ≤ µ0 (Uj )
G j=1 G j=1 j=1
P∞
and therefore by definition of µ0 , we conclude that µ0 (U ) ≤ j=1 µ0 (Uj ).

As a consequence of the lemma and basic measure theory (e.g. [Fol99, 1.10]),
we learn that µ∗ is an outer measure. Now, we show that every open set is µ∗ -
measurable, that is, let U be open and we want to show that whenever E ⊆ X is
such that µ∗ (E) < ∞, we will have µ∗ (E) ≥ µ∗ (E ∩ U ) + µ∗ (E ∩ U c ). Fix  > 0.
If RE happens to be open, then we can find f1 ∈ Cc (G) with f1 ≺ E ∩ U such
that G f1 ωR > µ0 (E ∩ U ) − , and similarly find f2 ∈ Cc (G) with f2 ≺ E − supp(f1 )
such that G f2 ω > µ0 (E − supp(f1 )) − . Note f1 + f2 ≺ E and so
Z
µ∗ (E) = µ0 (E) ≥ (f 1 + f2 ) ω > µ0 (E ∩ U ) + µ0 (E − supp(f1 )) − 2
G
= µ∗ (E ∩ U ) + µ∗ (E − supp(f1 )) − 2 ≥ µ∗ (E ∩ U ) + µ∗ (E ∩ U c ) − 2
For E not necessarily open, we find open V such that V ⊇ E and µ∗ (V ) = µ0 (V ) <
µ∗ (E) + . Then
µ∗ (E) +  > µ∗ (V ) ≥ µ∗ (V ∩ U ) + µ∗ (V ∩ U c ) ≥ µ∗ (E ∩ U ) + µ∗ (E ∩ U c )
and so the desired inequality also holds.
12

Therefore, the collection of µ∗ -measurable sets will contain B(G), and we can
define µ := µ∗ |B(G) as a Borel measure.
Step 2: We check that µ is invariant. Let E ∈ B(G) and g ∈ G. Note that left
multiplication by g is a homeomorphism, so gE ∈ B(G). By definition,
µ (gE) = inf {µ0 (U ) | U ⊇ gE, U open} = inf {µ0 (gU ) | U ⊇ E, U open}
But
Z 
µ0 (gU ) = sup f ω | f ∈ Cc (G), f ≺ gU
G
Z 
∗
= sup f (Lg−1 ω) | f ∈ Cc (G), f ≺ U = µ0 (U )
G
where L∗g−1 denotes the pullback by the diffeomorphism Lg−1 that is left multi-
plication by g −1 , and where we used the left-invariance of ω in the last equality.
Therefore µ(gE) = µ(E) as desired. 
As a step towards increasing generality, let G be a Lie group and H a closed
subgroup of G. The left coset space G/H can be equipped with a smooth manifold
structure such that the quotient map π : G → G/H is smooth [Lee12, 21.17]. Under
this structure, G/H can be acted on by G via
g1 g2 = g1 g2
and this action is smooth.
Proposition 2.1.3. There is an invariant Borel measure for G/H equipped with
the above action by G.
Proof. Equip G with the invariant Borel measure µ. G/H here is equipped with
the Borel σ-algebra. The map π : G → G/H is smooth, thus continuous, and hence
measurable. Therefore, the pushforward measure π∗ (µ) is a Borel measure on G/H.
It is invariant because for S ∈ B(G/H), we have gS ∈ B(G/H) because the action
of g on G/H is a homeomorphism, and
(π∗ (µ))(gS) = µ(π −1 (gS)) = µ(g(π −1 (S))) = µ(π −1 (S)) = (π∗ (µ))(S)
as desired. 
The above consideration involving G/H is actually a lot more general than it
might first appear. A homogeneous G-space (or simply homogeneous space)
is a smooth manifold endowed with a transitive smooth action by the Lie group
G. Clearly, G/H is an example of a homogeneous G-space. The following theorem
tells us that in fact all homogeneous spaces look like this.
Theorem 2.1.4. Let G be a Lie group and M a homogeneous G-space. Pick any
point p ∈ M . Then the stabilizer of p, denoted Stab(p) is a closed subgroup of G
and the map F : G/Stab(p) → M defined by F (g) = gp is a diffeomorphism.
Proof. See [Lee12, 21.28]. 
Corollary 2.1.5. A homogeneous G-space can be equipped with an invariant Borel
measure.
Proof. Pick any p ∈ M and consider the pushforward of the invariant Borel measure
of G/Stab(p) using the map F as defined in theorem 2.1.4. 
 13

Finally, the following theorem is useful if we start with a set that might not yet
be equipped with a smooth manifold structure.

Theorem 2.1.6. Let T be a set, endowed with a transitive action by a Lie group
G such that for some p ∈ T , we know that the stabilizer subgroup Stab(p) is closed
in G. Then T can be made into a smooth manifold such that it is a homogeneous
G-space.

Proof. See [Lee12, 21.20]. 

2.2. Grassmannians. The above discussion can be applied to two examples of

particular interest to integral geometry. The first is Gr(n, k), the set of all k-
dimensional subspaces of Rn , acted transitively upon by O(n) (k here is some fixed
integer satisfying 0 ≤ k ≤ n). If we arbitrarily pick an element L0 ∈ Gr(n, k), then
Stab(L0 ) is some copy of O(k) × O(n − k) inside O(n) which can be shown to be a
closed subgroup. For example, if we view O(n) as the set of n × n orthogonal matri-
ces with respect to the standard basis (1 , . . . , n ) and select L0 = span(1 , . . . , k ),
then Stab(L0 ) is the set of matrices of the form
 
A 0
0 D

where A and D are orthogonal matrices of sizes k×k and (n−k)×(n−k) respectively.
We shall abuse notation and understand O(k) × O(n − k) as a specific subgroup of
O(n). The smooth manifold structure (which includes the topology) of Gr(n, k) is
then better understood as
O(n)
Gr(n, k) '
O(k) × O(n − k)

and has an O(n)-invariant Borel measure.

Let θn denote a O(n)-invariant Borel measure on O(n) itself, which exists by the-
orem 2.1.2. The corresponding O(n)-invariant Borel measure on Gr(n, k), denoted6
by γn,k , is obtained by (two) pushforwards and concretely is given by

γn,k (A) = θn ({g ∈ O(n) | gL0 ∈ A})

for any A ∈ B(Gr(n, k)).

In turn, here is one way of defining some θn that is practical for our applica-
tions. Observe that there is a bijection between elements of O(n) and an ordered
orthonormal basis under the map g 7→ (e1 , . . . , en ) = (g(1 ), . . . , g(n )). We begin
by picking e1 ∈ S n−1 in accordance to the normalized spherical measure σn−1 (nor-
malized means σn−1 (S n−1 ) = 1). Now, e2 must be picked from the copy of S n−2
that lies in the hyperplane orthogonal to e1 , and we pick e2 in accordance to the
normalized spherical measure σn−2 . In general, we pick ei uniformly from the copy
of S n−i that lies in the (n − i + 1)-dimensional plane orthogonal to e1 , . . . , ei−1 . In

6We chose the symbol γ because “gamma” and “Grassmannians” both start with “g”. Likewise,
hopefully the symbol θ reminds one of angles and thus serves as a mnemonic for the measure on
O(n).
14

other words, we can define θn via

Z
f (g) θn (dg)
O(n)
Z Z Z
= ... f (i 7→ ei ) σ0 (den ) . . . σn−2 (de2 ) σn−1 (de1 )
S n−1 S n−2 ⊥{e1 } S 0 ⊥{e1 ,...,en−1 }

for any measurable function f (and i 7→ ei denotes the unique element of O(n)
that maps i to ei ).
From the above description of θn , we have γn,k (Gr(n, k)) = θn (O(n)) = 1, so
for each fixed n and k, Gr(n, k) is already a probability space. No normalization is
needed.
We will need the following lemma in the future.
Lemma 2.2.1. Let f be a measurable function on Gr(n, k). Then
Z Z
f (L) γn,k (dL) = f (L⊥ ) γn,n−k (dL)
Gr(n,k) Gr(n,n−k)

Proof. Pick any L0 ∈ Gr(n, k). For a subset A ⊆ Gr(n, k), we let A⊥ ⊆ Gr(n, n−k)
denote
A⊥ := L⊥ | L ∈ A


Observe that
γn,k (A) = θn ({g ∈ O(n) | gL0 ∈ A}) = θn ( g ∈ O(n) | gL⊥ ⊥
) = γn,n−k (A⊥ )

0 ∈A
from which the desired conclusion follows. 
2.3. k-planes. A similar discussion can be made for Aff(n, k), the set of all k-
dimensional planes in Rn , acted upon transitively by E(n). Recall that E(n) can
be written as the semi-direct product T (n) o O(n), where T (n) is the translation
group and O(n) acts on T (n) ' Rn in the usual way (for a review of Lie group
structure under semi-direct products, see [Lee12, 7.32]). As a matrix group, E(n)
can be realized as matrices of the form
 
A11 . . . A1n b1
 .. .. .. .. 
 .
 . . . 

 An1 . . . Ann bn 
0 ... 0 1
where A is an orthogonal matrix (standard basis chosen here) and b is a vector
(again, standard basis). Pick an element P0 ∈ Aff(n, k). Then Stab(P0 ) is some
copy of (T (k)×{e})o(O(k)×O(n−k)) inside E(n) which can be shown to be a closed
subgroup. For example, without loss of generality, suppose P0 = span(1 , . . . , k ).
Then Stab(P0 ) consists of matrices of the form
 0 
A 0 b
 0 A00 0 
0 0 1
where A0 is orthogonal of size k × k, A00 is orthogonal of size (n − k) × (n − k) and
b is a vector of length k. We will abuse notation and understand (T (k) × {e}) o
(O(k) × O(n − k)) as a subgroup of E(n). Then, as smooth manifolds,
E(n)
Aff(n, k) '
(T (k) × {e}) o (O(k) × O(n − k))
 15

and has an invariant Borel measure.

There is a natural candidate for an invariant Borel measure on E(n). First, we
perform a verification that B(E(n)) = B(T (n)) ⊗ B(O(n)). Indeed, the smooth
manifold structure of semi-direct product is taken to be that of the Cartesian prod-
uct, in which case the topology is the product topology. Now T (n) and O(n) are
Lie groups, hence second-countable and thus separable; they are also metrizable
2
(they inherit the metric from Rn and Rn respectively). The result follows from
this standard lemma from measure theory.
QN
Lemma 2.3.1. If X1 , . . . , XN are metric spaces and X is j=1 Xj equipped with
NN
the product topology, then j=1 B(Xj ) ⊆ B(X). If furthermore X1 , . . . , XN are
NN
separable, then j=1 B(Xj ) = B(X).
Proof. See any text in measure theory, for example [Fol99, 1.5]. 
Let λn denote the n-dimensional Lebesgue measure, which we restrict to Borel
sets but continue to use the same symbol as a mild abuse of notation, and recall θn
from section 2.2. Since we have B(E(n)) = B(T (n)) ⊗ B(O(n)), it makes sense for
us to talk about the product measure λn × θn as a Borel measure on E(n). One can
check that this is an E(n)-invariant measure. Under the pushforwards, we obtain
an E(n)-invariant measure on Aff(n, k), which we denote7 by αn,k .
One can work through the details of the pushforwards and obtain the following
practical interpretation of αn,k : choosing a k-plane can be seen as choosing an
element of Gr(n, k) in accordance to γn,k and then translating the subspace along a
direction in its orthogonal complement chosen in accordance to λk . In other words,
we can understand αn,k as
Z Z Z
f (P ) αn,k (dP ) = f (L + y) λk (dy) γn,k (dL)
Aff(n,k) Gr(n,k) L⊥

for any measurable function f .

2.4. Example - Buffon’s Needle Problem in Rn . As a nice opportunity to
illustrate the ideas of this chapter, let us finish off the Buffon’s Needle Problem
that was unsettled in the introduction.
Example 2.4.1 (Buffon’s Needle Problem in Rn ). Let there be a needle of length
l fixed in Rn , n ≥ 2. We have to choose a probability measure on systems of
parallel hyperplanes spaced a unit apart, and hopefully obtain a relation between
the expected number of intersections and l.
We have already discussed that because the needle length is invariant under
E(n), we had better choose a probability measure that is E(n)-invariant. The set
T containing all systems of hyperplanes is equipped with the action by E(n), and
very similar to the discussion of k-planes, we can obtain an E(n)-invariant Borel
measure α en,n−1 where
Z Z Z 1
f (P ) α
en,n−1 (dP ) = f (L + y) λ1 (dy) γn,n−1 (dL)
T Gr(n,n−1) 0

for any measurable function f . The inner integral ranges from 0 to 1 because
translating the system of hyperplanes along the normal direction will lead to the
7Again, “lambda” and “Lebesgue” both start with the same letter, as do “alpha” and “affine”.
16

same configuration every integer distance. Note that αen,n−1 is already normalized
to be a probability measure because take f (P ) ≡ 1 and we get
en,n−1 (T ) = λ1 ([0, 1]) · γn,n−1 (Gr(n, n − 1)) = 1
α
Let X : T → R ∪ {±∞} denote the number of intersections, which can be shown
to be a random variable, i.e. measurable. First assume 0 < l < 1. Then
Z
E [X] = X(P ) α
en,n−1 (dP )
T
Z Z 1
= X(L + y) λ1 (dy) γn,n−1 (dL)
Gr(n,n−1) 0

Now, the inner integral is simply the length of the needle after projected onto
L⊥ . Introducing the notation S|L to denote the orthogonal projection of a subset
S ⊆ Rn onto a linear subspace L, we want to compute
Z
length(needle|L⊥ ) γn,n−1 (dL)
Gr(n,n−1)
Z
= length(needle|L) γn,1 (dL)
Gr(n,1)

where we have applied lemma 2.2.1.

Now we have to do actual computations. Pick L0 ∈ Gr(n, 1) to be the line
spanned by the first standard basis vector 1 . For a subset A ⊆ Gr(n, 1), note that
γn,1 (A) = θn ({g ∈ O(n) | gL0 ∈ A})
( )!
[
= θn g ∈ O(n) | g1 ∈ (L ∩ S n−1 )
L∈A

So, by our definition of θn , the integral becomes

Z
length(needle|span(u)) σn−1 (du)
S n−1
Z
=l |v · u| σn−1 (du)
S n−1
2 · vol(B n−1 )
= l
area(S n−1 )
where v is a unit vector pointing along the direction of the needle and B n−1 is
the closed unit sphere in Rn−1 . The last equality is by lemma 2.4.2 below, with a
sketched proof for those who have not seen this fact from calculus. Note that σn−1
here is the normalized spherical measure, hence the differing factor with lemma
2.4.2.
Substituting general expressions for volume and surface area of spheres, we con-
clude that
2 Γ( n2 + 1)
E [X] = √ l
n π Γ( n2 + 21 )
For example, for n = 2, 3, 4, we respectively get the proportionality constant to be
2/π, 1/2 and 4/(3π). 
 17

Lemma 2.4.2. Let v be a fixed unit vector. Then

Z
|v · n| dS = 2 · vol(B n−1 )
S n−1
where dS is the usual surface element and n is the outward unit normal vector.
Proof. (Sketch) A full proof can be found in [KR97, 5.5.1]. Chop up the surface
into many small regions Ai with surface area S(Ai ) and we have the approximation
Z X
|v · n| dS ≈ |v · ni | S(Ai )
S n−1 i
Observe that |v · ni | S(Ai ) is approximately the area of the shadow when Ai is
orthogonally projected onto the hyperplane v ⊥ . On orthogonal projection, each
hemisphere as separated by v ⊥ covers the solid unit ball in v ⊥ once. Thus the
above sum is approximately equal to 2 · vol(B n−1 ). The approximations become
equality in the limit. 
3. Notions of Size
Mathematics is the art of giving the same name to different things.
– Henri Poincaré

We are ready to see how expectations can be used to define reasonable notions
of size. Actually the Buffon’s Needle Problem of example 2.4.1 gave us a glimpse
of this idea – the length of a needle can be computed (up to scaling factor) as the
expectation of a random variable. However, the example is not very satisfactory
on two accounts: the length of a line segment is not a difficult concept to grapple
with, and we want to avoid having to search from scratch a random variable for
each reasonable notion of size that one can think of. Fortunately, this chapter will
explain how integral geometry provides an elegant and unified way to think about
all reasonable notions of size, many of which are difficult to grasp by classical
approaches.
As the Banach-Tarski paradox shows, attempting to assign sizes to every subset
of Rn is a tricky business. Our theory will focus only on Kn , the collection of
all compact convex subsets of Rn . This collection is wide enough to include some
examples of smooth manifold (or rather, the boundary of the subset is a smooth
manifold), polyhedrons, but also other less well-behaved examples. Yet, it is narrow
enough to produce an elegant theory. We shall also remark that once we completed
a discussion for Kn , it is possible to extend our size functions to the collection of
polyconvex sets (i.e. the collection consisting of finite unions of compact convex
sets) by a direct application of an extension theorem by Groemer. However, we will
not pursue this discussion here – a good reference on Groemer’s extension theorem
will be [KR97, Chapter2]. The collection of polyconvex sets practically covers all
geometrical objects one would deal with in real life8; even this ‘o’ is really a finite
union of convex pixels9. We merely wish to point out that the generality of this
theory, though limited to Kn , is not something to be underestimated.
3.1. “Reasonable” notion of size. A notion of size will be a map φ : Kn → R,
but we should not allow every possible such mapping. There will be some properties
we hope φ should satisfy before it will be convincing to call φ a notion of size. Below,
three such properties are described.
Definition 3.1.1. The map φ : Kn → R is said to be a valuation if φ(∅) = 0 and
φ(K ∪ L) = φ(K) + φ(L) − φ(K ∩ L)
whenever K, L ∈ K are such that K ∪ L ∈ Kn .
n

Definition 3.1.2. The map φ : Kn → R is said to be invariant if φ(gK) = φ(K)

for all K ∈ Kn and g ∈ E(n), where gK := {g(x)|x ∈ K}.
8That said, I am not completely sure what mathematicians will regard as “real life”.
9In some sense, everything we encounter in real life is a finite union of atoms, but I will leave
it to the physicists to decide if we are permitted to talk about atoms as “convex”.
 19

We will want φ to be an invariant valuation. Before describing the third property,

it will be useful to recall what is known as the Hausdorff metric.
Definition 3.1.3. For K, L ∈ Kn , define the Hausdorff distance between them
to be
   
δ(K, L) := max sup d(k, L), sup d(K, l) = max sup inf d(k, l), sup inf d(k, l)
k∈K l∈L k∈K l∈L l∈L k∈K
n
where d refers to the Euclidean metric in R .
Note that δ is not yet shown to be a metric. Before doing that, here is a lemma
that provides a more intuitive way of understanding δ. The statement of the lemma
requires us to first understand the -fattening of an object in Kn , where  ≥ 0. We
define the -fattening of ∅ to be ∅, and for non-empty K ∈ Kn ,
K+ := {x ∈ Rn | d(K, x) ≤ }
Recall that for K, L ∈ Kn , their Minkowski sum is given by
K + L := {k + l | k ∈ K, l ∈ L}
Thus, we can also write K+ = K + B n , where B n is the solid unit sphere of
radius . Note that K+ is bounded, closed (from continuity of d(K, ·)) and convex
(given x, y ∈ K+ , decompose x, y using the Minkowski sum and then use the fact
that both K and B n are convex), whence K+ ∈ Kn .
Lemma 3.1.4. Let K, L ∈ Kn . Then δ(K, L) ≤  if and only if K ⊆ L+ and
L ⊆ K+ . Hence, we may equivalently define
δ(K, L) = inf { ≥ 0 | K ⊆ L+ and L ⊆ K+ }
and in fact, the above infimum is achieved.
Proof. If L ⊆ K+ , then every l ∈ L obeys d(K, l) ≤ . Similarly if K ⊆ L+ , then
every k ∈ K obeys d(k, L) ≤ . Therefore,
 
δ(K, L) = max sup d(k, L), sup d(K, l) ≤ 
k∈K l∈L

Conversely, we prove the contrapositive. WLOG suppose L ⊆ K+ is violated, so

there exists ˜l ∈ L such that d(K, ˜l) > . Then δ(K, L) ≥ d(K, ˜l) > .
Note that if K ⊆ L+ and L ⊆ K+ hold, then K ⊆ L+η and L ⊆ K+η also hold
for any η > . Thus
E := { ≥ 0 | K ⊆ L+ and L ⊆ K+ }
is either of the form (c, ∞) or [c, ∞) for some c ≥ 0. By the first half of the lemma,
δ(K, L) cannot be bigger or smaller than c, so δ(K, L) = c. Then again by the first
half of the lemma, c ∈ E so E = [c, ∞) and the infimum is achieved. 
Proposition 3.1.5. The Hausdorff distance δ is a metric and turns Kn into a
metric space.
Proof. First, δ maps into R because our sets are compact (in particular bounded).
δ is clearly symmetric, non-negative, and δ(K, K) = 0 for K ∈ Kn . If K, L ∈ Kn
are distinct, at least one of K − L or L − K will be non-empty; WLOG say we can
pick some ˜l ∈ L − K. Since K is compact (in particular closed), d(K, ˜l) > 0. Thus
δ(K, L) ≥ d(K, ˜l) > 0.
20

It remains to prove the triangle inequality. Let K, L, M ∈ Kn and  = δ(K, M ),

η = δ(L, M ). By lemma 3.1.4,
K ⊆ M+ and M ⊆ L+η
and so K ⊆ L+(+η) and similarly L ⊆ K+(+η) . Again by lemma 3.1.4, δ(K, L) ≤
 + η. 
Now we may state the final property that φ should have.
Definition 3.1.6. The map φ : Kn → R is simply said to be continuous if it is a
continuous map with the Hausdorff metric on Kn and the usual Euclidean metric
on R.
In summary, we shall be interested in studying the collection of all continuous
invariant valuations on Kn , which we shall denote Valn . Here is an easy property
of Valn that already provides a lot of information about its structure.
Proposition 3.1.7. Valn is a real vector space under the usual definition of how
to add functions and perform scalar multiplication.
Proof. It is clear that for φ, ψ ∈ Valn and c ∈ R,
• φ, ψ are valuations ⇒ φ + ψ and cφ are valuations
• φ, ψ are invariant ⇒ φ + ψ and cφ are invariant
• φ, ψ are continuous ⇒ φ + ψ and cφ are continuous
Therefore we have φ + ψ ∈ Valn and cφ ∈ Valn . 
Example 3.1.8. One natural candidate element of Valn would be the n-dimensional
Lebesgue measure (restricted to Kn ), which we shall still denote by λn . It follows
from any standard exposition on Lebesgue measure that this is an invariant valu-
ation. What is believable intuitively but not so clear rigorously is continuity. This
has been established in [Bee74]. 
Example 3.1.9. It is not obvious how we might define the perimeter of a compact
convex subset in R2 (the boundary may behave rather wildly!). However, if we
temporarily restrict ourselves to P2 ⊆ K2 defined to be the collection of all compact
convex polygons (including degenerate ones which are simply line segments), then
we have a clear way of defining what we mean by perimeter10. We leave it as an
exercise for the reader to show that the perimeter is a continuous invariant valuation
on P2 . Later, we will see how integral geometry allows us to easily produce a
continuous invariant valuation on K2 that is essentially the idea of perimeter (i.e.
the two valuations coincide when restricted to P2 ). A similar discussion can be
made for surface area of compact convex subsets in R3 . 
3.2. Intrinsic Volumes. Here is one natural way to measure the size of an object
K in Kn which will indeed turn out to be a continuous invariant valuation. The
idea is to take a random k-subspace L of Rn (1 ≤ k ≤ n), look at the orthogonal
projection K|L of K onto that plane, and compute the expected k-dimensional
Lebesgue measure of the resulting shadow. We want our size function to be in-
variant, which means the probability measure we impose on Gr(n, k) should be the
O(n)-invariant measure γn,k .

10For line segments, we have to “double-count” the length.

 21

Definition 3.2.1. For n ≥ 1, the intrinsic volume functions Vn,0 , . . . , Vn,n : Kn →

R are defined as follows. For 1 ≤ k ≤ n,
Z
Vn,k (K) := λk (K|L) γn,k (dL)
Gr(n,k)

and for Vn,0 , we define Vn,0 (K) to be 1 whenever K non-empty and 0 otherwise.
A couple of remarks about the definition needs to be made. First, we handled
the Vn,0 case separately but if we wish, we could have thought about “orthogonal
projection” onto the origin and regard λ0 as the counting measure. Second, one
can see that K|L is indeed a compact subset of L and hence λk (K|L) makes sense
and is finite; moreover, one should technically check that λk (K|·) is a measurable
function, but here we will be content in saying that the Borel σ-algebra is fine
enough to handle the behavior of compact convex K in this aspect. Third, Vn,n
agrees with λn .
Proposition 3.2.2. The intrinsic volume functions are continuous invariant valu-
ations.
Proof. The claim is clearly true for Vn,0 , so let us ignore that case.
Invariance under E(n) follows from the O(n)-invariance of γn,k , the translational
invariance of λk in the subspace L, and the fact that translating K in the direction
orthogonal to L does not change K|L.
Next, we show continuity of Vn,k . Suppose K1 , K2 , . . . , K ∈ Kn are such that
Ki → K. Note that there is a sufficiently giant closed ball B that contains all of
K1 , K2 , . . . , K. For fixed L, it is easy to see from lemma 3.1.4 that (Ki |L) → (K|L).
Since λk is continuous, λk (Ki |L) → λk (K|L). Finally, observe that λk (Ki |L) ≤
λk (B|L) for all i and so we may apply the dominated convergence theorem to
conclude that
Z Z
Vn,k (Ki ) = λk (Ki |L) γn,k (dL) → λk (K|L) γn,k (dL) = Vn,k (K)
Gr(n,k) Gr(n,k)

Finally, we show that Vn,k is a valuation. Suppose K1 , K2 ∈ Kn are such that

K1 ∪ K2 is also convex. Note that for any L, we have
(K1 ∪ K2 )|L = (K1 |L) ∪ (K2 |L)
and that this is convex. A priori, we also have
(K1 ∩ K2 )|L ⊆ (K1 |L) ∩ (K2 |L)
but we must have equality. Indeed, pick x ∈ (K1 |L) ∩ (K2 |L). Define x + L⊥ to be
the translated copy of L⊥ that passes through x. Let
Xi := {k ∈ Ki | k projects to x} = Ki ∩ (x + L⊥ )
Note that X1 , X2 and X1 ∪ X2 are non-empty compact convex. By the lemma
following this proposition, X1 ∩ X2 is non-empty so x ∈ (K1 ∩ K2 )|L. Having
proven the identities
(K1 ∪ K2 )|L = (K1 |L) ∪ (K2 |L)
(K1 ∩ K2 )|L = (K1 |L) ∩ (K2 |L)
We conclude using the fact that λk is a valuation and the linearity of integration. 
22

Lemma 3.2.3. Suppose K, L ∈ Kn non-empty are such that K ∪ L ∈ Kn . Then

K ∩ L ∈ Kn is non-empty.
Proof. If L ⊆ K, we are done. Otherwise, pick any l ∈ L−K. Since K is closed and
convex, it follows from theory of Hilbert spaces that there exists a unique k ∈ K
that is the closest point of K to l. Consider the line segment joining k and l, which
must lie in K ∪ L by convexity of K ∪ L. By construction, k must be the only point
on the line segment that lies in K, so all other points of the line segment must lie
in L. Since L is closed, we conclude k ∈ L and so k ∈ K ∩ L. 
There is an alternative idea that is first explored by Crofton, which is to pick a
random (n−k)-dimensional hyperplane and ask whether it will intersect the convex
body. Intuitively, the larger the body, the more “likely” an intersection will occur.
It turns out that this idea leads back to the same intrinsic volume functions. In
the proposition, note that Vn,0 has be used to detect whether an intersection has
occurred or not.
Proposition 3.2.4. (Crofton formula) For 1 ≤ k ≤ n, we have
Z
Vn,k (K) = Vn,0 (K ∩ P ) αn,n−k (dP )
Aff(n,n−k)

Before proceeding with the proof, here is an obligatory remark that one should
technically check that Vn,0 (K ∩ ·) is a measurable function for the integral to make
sense, but we will again leave the issue aside.
Proof. We have
Z
χ(K ∩ P ) αn,n−k (dP )
Aff(n,n−k)
Z Z
= χ(K ∩ (L + y)) λk (dy) γn,n−k (dL)
Gr(n,n−k) L⊥

Now, for a fixed L ∈ Gr(n, n − k), as y varies in L⊥ , K ∩ (L + y) is non-empty if

and only if y ∈ K|L⊥ . Therefore, the above expression simplifies to
Z
λk (K|L⊥ ) γn,n−k (dL)
Gr(n,n−k)
Finally, we recall proposition 2.2.1 and conclude that this expression is equal to
Z
λk (K|L) γn,k (dL)
Gr(n,k)

which is Vn,k (K). 

3.3. Hadwiger’s Theorem. Recall from proposition 3.1.7 that Valn is a real vec-
tor space. The key importance of intrinsic volume functions lies in the following
theorem, which tells us that the intrinsic volume functions that we have defined
through integral geometric language are really all that is needed to understand the
collection of continuous invariant valuations.
Theorem 3.3.1. (Hadwiger) The intrinsic volume functions Vn,0 , . . . , Vn,n are a
basis for Valn . In particular, dim(Valn ) = n + 1.
Proof. A typical proof is long and will derail us form the discussion. We refer
interested readers to [Kla95]. 
 23

Actually, it is not difficult to see that Vn,0 , . . . , Vn,n are linearly independent. A
function φ : Kn → R is said to be homogeneous of degree i if φ(tK) = ti φ(K) where
tK := {tx | x ∈ K}. Since the k-dimensional Lebesgue measure is homogeneous of
degree k, it follows from the definition of the intrinsic volume functions that Vn,k
is homogeneous of degree k. If a0 , . . . , an are scalars such that
a0 Vn,0 + · · · + an Vn,n ≡ 0
then we simply evaluate the above expression on tBn and obtain
a0 Vn,0 (Bn )t0 + · · · + an Vn,n (Bn )tn = 0 for all t ≥ 0
which means the above polynomial in t is the zero polynomial. Since Vn,i (Bn ) 6= 0
for all i, we conclude that a0 = · · · = an = 0. Using a similar idea, we can deduce
the following corollary from theorem 3.3.1.
Corollary 3.3.2. If φ is a continuous invariant valuation that is homogeneous of
degree i, then it must be that φ = cVn,i for some real constant c.
Proof. By theorem 3.3.1, we can write φ = a0 Vn,0 + · · · + an Vn,n for some scalars
a0 , . . . , an . Evaluate both sides of the expression on tBn and conclude as above
that all the scalars are zero except possibly ai = φ(Bn )/Vn,i (Bn ). 
Now, we are in a position to perform a normalization procedure that captures
the full meaning of the word “intrinsic” in the name “intrinsic volume”. Suppose
K ∈ Km . If n ≥ m, we may also consider Rm as a subset of Rn in the most natural
way and view K as an element of Kn . There is no guarantee that Vm,k (K) =
Vn,k (K). However, we can easily rescale our intrinsic volume functions to achieve
this desirable property, without affecting any of the theory that we have developed
thus far.
We proceed inductively. To prevent confusion, V will continue to denote the
intrinsic volumes as we have defined previously, and W will be used to denote the
rescaled versions of V . We want to define Wn,k for all n ≥ 1, 0 ≤ k ≤ n.
(1) For all n ≥ 1, we let Wn,0 = Vn,0 . Also set W1,1 = λ1 which is also V1,1 .
(2) Suppose Wn−1,k has been set for all 0 ≤ k ≤ n − 1. First set Wn,n = λn
which is also Vn,n . Next, consider a fixed 1 ≤ k ≤ n − 1 and we want
to define Wn,k . If K ∈ Kn−1 , then we can consider it as an element of
Kn and compute Vn,k (K). Thus Vn,k induces a k-homogeneous continuous
invariant valuation on Kn−1 . By corollary 3.3.2, we can write this induced
map as cVn−1,k for some c 6= 0, which is also of the form c0 Wn−1,k for some
c0 6= 0. Set Wn,k = c10 Vn,k .
By construction, Wn,k has the desired property that Wm,k (K) = Wn,k (K) when-
ever n ≥ m and K ∈ Km (and the bonus property that Wn,n coincides with
Lebesgue measure). We will call the Wn,k the rescaled intrinsic volume func-
tions. Now the intrinsic volumes of K are indeed “intrinsic” to K itself, in the
sense that it does not depend on the dimension of the ambient space.
3.4. Steiner’s formula. We now consider a formula due to Steiner that will give
an alternative interpretation of the intrinsic volume functions. For example, we will
see (up to scaling factors) how W2,1 can be thought of as the perimeter function, of
how W3,2 can be thought of as the surface area function, generalized to all compact
convex subsets. The correct scaling factor can be easily obtained by performing
computation on say B n and is an unimportant detail theoretically.
24

ε ε

ε ε

ε ε

λ2 (T+ε ) = area(T ) + ε · perimeter(T ) + ε2 · π

Figure 3.1.

ε
ε
ε

total edge length(C) 3 4π

λ3 (C+ε ) = volume(C)+ε·surface area(C)+ε2 ·π· +ε ·
4 3

Figure 3.2.

Steiner’s formula aims to describe how the n-dimensional Lebesgue measure

of a fixed compact convex subset of Rn changes as it undergoes -fattening. It
turns out that λn (K+ ), as a function of , will be a ≤ n-degree polynomial with
coefficients involving the intrinsic volumes Wn,0 (K), . . . , Wn,n (K). Figures 3.1 and
3.2 illustrate that we indeed obtain a polynomial in  for the fixed triangle T ⊆ R2
and the fixed cube C ⊆ R3 , and will be helpful for understanding the various
definitions made in the proof below. (Note: the polynomial would be different if
we had considered T to be in R3 - we would get a degree 3 polynomial with the
constant term being zero.)
 25

To begin the investigation, we will first obtain the formula for λn (P+ ) for P a
polytope.
Definition 3.4.1. A polytope in Rn is a bounded non-empty subset which can be
represented as the intersection of finitely many closed halfspaces.
Note that a polytope is necessarily in Kn (and it can very well be a “lower-
dimensional” object - because two closed halfspaces can intersect to give a hyper-
plane and further intersections after that will give a “lower-dimensional” polytope).
In fact, the following proposition suggests why it might be useful to first study the
collection Pn of polytopes.
Proposition 3.4.2. Under the Hausdorff metric, Pn is a dense subset of Kn .
Proof. Fix K ∈ Kn and  > 0. We wish to show the existence of P ∈ Pn for which
δ(P, K) ≤ . For each point k ∈ K, we consider the open ball ballk () centered at
k of radius . Then [
ballk ()
k∈K
is an S
open cover of K. By compactness of K, there exists k1 , . . . , kw such that
w
K ⊆ i=1 ballki (). Set P to be the convex hull of k1 , . . . , kw . Then P ∈ Pn and
because K is convex, we have P ⊆ K ⊆ K+ . By construction of P , we have
K ⊆ P+ . Therefore by lemma 3.1.4, we have δ(P, K) ≤ . 
Recall that a supporting hyperplane of P ∈ Pn is a hyperplane H such that
H ∩ P 6= ∅ and P lies entirely in one of the two closed halfspaces defined by H.
In fact, H ∩ P is again a polytope and is called an k-face if its dimension is k
(0 ≤ k ≤ n − 1). We also consider P to be a n-face of itself. For 0 ≤ k ≤ n, we
let Fk be theSncollection of all k-faces of P and denote the collection of all faces of
P by F := k=0 Fk . For a face F ∈ F, we define its relative interior relint(F )
to be all elements of F that do not belong to any face of strictly lower dimension.
Observe that P can be written as the disjoint union
a
P = relint(F )
F ∈F
In fact, we can decompose the fattened P+ into a disjoint union as well. From
theory of Hilbert space, to each point x ∈ Rn , we can associate a unique closest
point in P ; let proj : Rn → Rn denote this map (the image will be in P ). By the
partition above, we end up with a partition of Rn depending on where proj(x) lies.
Thus we may write
a
P+ = (P+ ∩ proj−1 (relint(F )))
F ∈F
Now we have to find a way to compute the size of each piece. For 0 ≤ k ≤ n − 1
and F ∈ Fk we define the normal cone of F , denoted N (F ), as follows: pick any
x ∈ relint(F ), and consider the closed convex cone consisting of the zero vector and
all (not necessarily unit) outer normal vectors of supporting hyperplanes at x. It
is easy to check that this definition does not depend on the choice of x. Finally, we
observe that
λn (P+ ∩ proj−1 (relint(F )))
= λn−k (N (F ) ∩ B n ) · λk (F )
= n−k · λn−k (N (F ) ∩ B n ) · λk (F )
26

and of course for the separate case k = n, we have λn (P+ ∩ proj−1 (relint(P ))) =
λn (P ).
Therefore, if we set, for 0 ≤ k ≤ n − 1,
X
(3.1) W̃n,k (P ) := λn−k (N (F ) ∩ Bn ) · λk (F )
F ∈Fk

and W̃n,n (P ) := λn (P ), then we have the following theorem.

Theorem 3.4.3. (Steiner’s formula for polytopes) For a polytope P ∈ Pn , λn (P+ )
as a function of  is a ≤ n-degree polynomial. In fact, with W̃n,0 , . . . , W̃n,n : Pn → R
defined as above,
n
X
λn (P+ ) = n−k · W̃n,k (P )
k=0

Proof. The proof is given by the discussion above. 

Now, we extend the discussion from Pn to Kn .

Theorem 3.4.4. (Steiner’s formula) There are functions W̃n,0 , . . . , W̃n,n : Kn → R
such that for any K ∈ Kn , we have
n
X
λn (K+ ) = n−k · W̃n,k (K)
k=0

Proof. First, observe that for any polytope P ∈ Pn , we have the following system
of linear expressions from theorem 3.4.3:
    
λn (P+1 ) 1 1 ... 1 W̃n,0 (P )
 λn (P+2 )   2n 2n−1 . . . 1  W̃n,1 (P ) 
 =  ..
    
.. .. .. .  ..
. ..  
 
 .   . . . 
λn (P+n ) nn nn−1 . . . 1 W̃n,n (P )
Inverting the Vandermonde matrix, we can express
n
X
W̃n,k (P ) = ckj · λn (P+j )
j=0

where ckj are entries of the inverted matrix. Now for any K ∈ Kn , we simply define
n
X
(3.2) W̃n,k (K) := ckj · λn (K+j )
j=0

We already noted that λn is continuous on Kn (example 3.1.8), but in fact for a fixed
 ≥ 0, the mapping K 7→ λn (K+ ) is continuous as well, since it is the composition
of continuous maps11 K 7→ K+ 7→ λn (K+ ). Therefore, the W̃n,k , defined by
equation (3.2), are continuous. Since we have Steiner’s formula for polytopes and
P is dense in K (proposition 3.4.3), we conclude that Steiner’s formula holds for
each K ∈ Kn by choosing a sequence of polytopes converging to K. 

11Continuity of the first map follows from the fact that K ⊆ L

+η implies K+ ⊆ (L+η )+ =
L+(η+) = (L+ )+η and an application of lemma 3.1.4.
 27

From the definition by equation (3.2), we see that W̃n,k is invariant under E(n),
and we already saw that it is continuous. In fact, W̃n,k is a valuation. This follows
directly from our previous observation that λn is a valuation (example 3.1.8) and
the following lemma (lemma 3.4.5), because from the lemma we know
n
X
W̃n,k (K ∪ L) = ckj · λn ((K ∪ L)+j )
j=0
Xn
= ckj · λn (K+j ∪ L+j )
j=0
Xn n
X n
X
= ckj · λn (K+j ) + ckj · λn (L+j ) − ckj · λn (K+j ∩ L+j )
j=0 j=0 j=0
Xn Xn Xn
= ckj · λn (K+j ) + ckj · λn (L+j ) − ckj · λn ((K ∩ L)+j )
j=0 j=0 j=0

= W̃n,k (K) + W̃n,k (L) − W̃n,k (K ∩ L)

whenever K, L ∈ Kn are such that K ∪ L ∈ Kn .
Lemma 3.4.5. Suppose K, L ∈ K are such that K ∪ L ∈ K. Then
(K ∪ L)+ = K+ ∪ L+
(K ∩ L)+ = K+ ∩ L+
Proof. x ∈ (K ∪ L)+ if and only if there exists some y ∈ K∪L such that d(x, y) ≤ ,
which is equivalent to either there being a y ∈ K or there being a y ∈ L such that
d(x, y) ≤ , which is equivalent to x ∈ K+ ∪ L+ .
One direction of the second equality is equally clear. Namely, if x ∈ (K ∩ L)+ ,
then there exists some y ∈ K ∩ L such that d(x, y) ≤ , so in particular d(K, x),
d(L, x) ≤ , whence x ∈ K+ ∩ L+ .To show the reverse inclusion, suppose x ∈
K+ ∩ L+ . Let k ∈ K and l ∈ L such that d(k, x) ≤  and d(l, x) ≤ . Let L be
the line segment joining k and l. Observe that K ∪ L is convex implies L ⊆ K ∪ L.
Also observe that k, l lie in the closed ball of radius  centered at x, so every point
of L is of distance ≤  from x. Therefore, it remains to show that some point on L
lies in K ∩ L. Indeed, L is connected and K ∩ L, L ∩ L are non-empty closed subsets
of L with (K ∩ L) ∪ (L ∩ L) = L, so K ∩ L and L ∩ L cannot be disjoint. 
From this discussion, we conclude that W̃n,k ∈ Valn . In fact, from the definition
by equation (3.1), we see that for any P ∈ Pn , we have
W̃n,k (tP ) = tk W̃n,k (P )
and by continuity with denseness of P , we conclude that W̃n,k is homogeneous of
degree k. Therefore, we have the following proposition.
Proposition 3.4.6. W̃n,k is equal to the intrinsic volume functions Wn,k up to a
scale factor.
Proof. This follows from corollary 3.3.2. 
For polytopes in R2 and R3 , it is easy to see that W̃2,1 and W̃3,2 are the perimeter
and surface area function respectively. Thus, we see how W2,1 and W3,2 are ways
28

to generalize these concepts to compact convex subsets of R2 and R3 respectively

(up to scaling factors). These generalizations are also natural ideas in the sense
that they are obtain by approximating with polytopes. The other W̃n,k ’s are also
not too difficult to understand concretely for polytopes, and provide a different
perspective of understanding our intrinsic volume functions Wn,k .
To sum up the rewards of this chapter, it is worth emphasizing that through the
lens of probability, we were able to generate the entire class of size functions on Kn
by the application of one single idea (be it looking at shadow size or intersection
counts). These size functions are not only generalizations of conventional notions
such as surface area and volume, but also apply to sets whose poor boundary
behavior typically exclude them from classical definitions.
4. Lengths on Manifolds
Everyone knows what a curve is, until he has studied enough mathemat-
ics to become confused through the countless number of possible exceptions,
– Felix Klein

The average shadow length when K ∈ Kn is projected onto a random element

of Gr(n, 1), which we denoted by Vn,1 , is also called the mean width. Proposition
3.2.4 due to Crofton tells us that Vn,1 can be equivalently understood as the measure
of hyperplanes that intersects K. For a 1-dimensional object in Kn , i.e. a line
segment, the rescaled version Wn,1 gives its length (because W1,1 agrees with λ1 ).
However, studying the length of line segments alone is not very interesting, and in
fact, studying the length of the wider class of piecewise smooth curves is a lot more
important, in the following sense.
Suppose M is a (smooth) Riemannian manifold where g is its Riemannian metric
(i.e. a smooth symmetric covariant 2-tensor field on M that is positive definite
at each point). Recall that M can be turned into a metric space as follows. If
γ : [a, b] → M is a piecewise smooth curve (from now all curves without any
adjectives are assumed to be piecewise smooth), define the length of γ to be
Z b
Lg (γ) := |γ 0 (t)|g dt
a
where γ 0 (t) lies in the tangent space at each point and so we may compute its norm
using g. We then define the distance between two points p, q ∈ M to be
dg (p, q) := inf Lg (γ)
γ

where γ ranges over all curves with endpoints p and q. It is a theorem that dg turns
M into a metric space whose metric topology coincides with the manifold topology
(for example, see [Lee12, 13.29]).
Although every smooth manifold M admits a Riemannian metric (see [Lee12,
13.3]), the idea is we want to start with just M alone and see if integral geometry
can be used to define length of curves on M . For example, in Rn , integral geometry
allowed us to reproduce the length of line segments and we have good intuitive
reason to believe that Crofton’s idea of looking at intersection with hyperplanes
will allow us to reproduce length of curves (proposition 4.1.1) - and indeed Crofton
himself proved this in R2 . Once we have length of curves on M , we can talk about
distances between points of M . This is a powerful alternative way to introduce
geometry (in a loose sense of the word) on M , rather than seeking for a Riemannian
metric. Note that this distance function might have nothing to do with the smooth
manifold topology of M .
The goal of this chapter is to lay the beginnings of this idea that is original to this
thesis. We will see how definitions of curve length permissible by our application
30

Figure 4.1.

of integral geometry include those obtainable by some Riemannian metric, but also
much more.
4.1. From Needles to Noodles. Before plunging into a general smooth mani-
fold M , let us first convince ourselves that a direct extension of Crofton’s formula
(proposition 3.2.4) from line segments to curves can be done. Intuitively, it says
that the length of a curve is proportional to the expected number of intersections
with randomly chosen hyperplanes (figure 4.1).
Proposition 4.1.1. (Crofton’s formula for curves) Fix n ≥ 1. For any curve γ, we
have Z
length(γ) = cn Nγ (P ) αn,n−1 (dP )
Aff(n,n−1)
where cn is a constant and Nγ : P → R ∪ {±∞} maps P to the number of inter-
sections between P and Im(γ).
Of course one should check that Nγ (·) is a measurable function. We will also
present a sketched proof rather than derail ourselves by fiddling with the details of
taking the limit.
Proof. (Sketch) We already have this result for γ a line segment. A key property
of Nγ that will give us the general result is additivity, namely, if γ : [a, b] → Rn is a
curve and a < z < b, then we can define γ1 := γ|[a,z] and γ2 := γ|[z,b] , and we have
N γ = N γ1 + N γ2
almost everywhere (there is some complication of double counting at the point
γ1 (z) = γ2 (z), hence the “almost everywhere”). Using additivity, we thus have the
result for γ a piecewise linear curve. Finally, a curve can be well-approximated by
a sequence of piecewise linear curves, from which the result follows. 
4.2. Extension to Manifolds. Let M be a smooth manifold and γ : [a, b] → M
be a curve. Our first goal is to find a way to define the length of γ using inspiration
from integral geometry. We would want to define the length as the expected number
of intersections with random choices of “planes”, and one way we can do this is to
use level sets of smooth functions on M .
 31

Definition 4.2.1. Let M be a smooth manifold and γ a curve. Whenever we have

a smooth function ω on M , denote by Nγ (ω) the number of intersections between
γ and the level set {x ∈ M |ω(x) = 0}. Now, if Ω is a set of smooth functions on M
and (Ω, Σ, µ) is a measure space such that Nγ : Ω → R ∪ {±∞} is measurable (i.e.
Σ needs to be sufficiently fine), then we may define
Z
Lµ (γ) := Nγ (ω) µ(dω)
Ω
We should remark that Nγ (ω), as the cardinality of the intersection set, may
well be infinite, and we have no reason to believe that this will occur with measure
zero. As such, Lµ (γ) may well be infinite.
For reasons that will be clear later, we will make a small change to what we mean
by Nγ (ω). We have the smooth function ω ◦ γ : [a, b] → M → R. Let us instead
define Nγ (ω) to be the number of transitions between ω ◦ γ ≤ 0 and ω ◦ γ > 0. This
notion is well-defined because ω ◦ γ is continuous and so the set of points satisfying
ω ◦ γ > 0 is an open subset of [a, b]. By the property of open subsets of R being a
countable disjoint union of open intervals, the same can be said for open subset of
[a, b] (except “open interval” here means relatively open, so it could be something
that includes the endpoints). Separating into the cases where the disjoint union
is countably infinite or finite, we respectively are able to compute an answer for
Nγ (ω).
Definition 4.2.2. Let M be a smooth manifold, γ a curve and Nγ (·) be defined
as above. If Ω is a set of smooth functions on M and (Ω, Σ, µ) is a measure space
such that Nγ : Ω → [−∞, ∞] is measurable, then we may define
Z
Lµ (γ) := Nγ (ω) µ(dω)
Ω
We will use definition 4.2.2 from now on. For a fixed M , a choice of (Ω, Σ, µ) for
which all Nγ are measurable (from now on implicitly assumed or verified) gives us
a definition of “length of curves”. In turn, a choice of (Ω, Σ, µ) gives us a distance
function
dµ (x, y) := inf Lµ (γ)
γ
where γ ranges over all curves with endpoints x and y, and the infimum is taken
to be ∞ if there is no such γ or if all values of Lµ (γ) happen to be infinite.
Although we allow (Ω, Σ, µ) to be incredibly general, the induced distance func-
tion has decent properties.
Proposition 4.2.3. Suppose (Ω, Σ, µ) is such that all pairs of points on M have
finite distance. Then dµ is a pseudometric (i.e. a metric except that one could have
dµ (x, y) = 0 for distinct x 6= y).
Proof. Let x, y, z be points of M . dµ (x, x) = 0 because we simply take γ to be the
curve γ([a, b]) = {x} for which Lµ (γ) = 0 since Nγ (ω) = 0 for all ω ∈ Ω. Next, it
is clear from the definition that dµ (x, y) = dµ (y, x). Lastly, the triangle inequality
holds because a path with endpoints x and y followed by a path with endpoints y
and z is in particular a path with endpoints x and z. 
The finiteness hypothesis of proposition 4.2.3 is not a serious obstacle. We can
define equivalence classes on M where x ∼ y if and only if dµ (x, y) < ∞. A similar
reasoning to the above proof shows that ∼ is indeed an equivalence relation. We
32

then have broken M into a disjoint union of so-called galaxies, in which dµ is a

pseudometric on each galaxy. Of course each galaxy may no longer be a manifold,
but we have already achieved some form of geometrical structure on each galaxy,
which is our real goal. One may further perform the metric identification on each
galaxy (i.e. x ∼ y if and only if dµ (x, y) = 0) and turn each galaxy into a metric
space.
Remark 4.2.4. The new definition for Nγ (ω) is crucial for us to obtain dµ (x, x) =
0. If we had stuck to definition 4.2.1, then Nγ (ω) need not be zero for the curve
γ([a, b]) = {x}, because the zero level curve of ω may very well contain x (so γ
intersects the level curve). In particular, consider the example where Ω contains
only the zero function. Then dµ (x, x) = 1 for all x ∈ M . 
Example 4.2.5. We can recover (up to a scaling factor) the Euclidean metric as
follows. Here M = Rn ,
Ω = x 7→ x · n + a | (n, a) ∈ S n−1 ×R


and we use the product (Borel) measure. The level curves are precisely hyperplanes.
There is a technical subtlety that Crofton’s formula for curves uses the notion of
Nγ in definition 4.2.1, but in this setting it can be shown that for any curve the
two definitions differ on hyperplanes that make up zero measure. 
The following is a generalization of the above example, and show that our integral
geometric framework is no less general than what Riemannian geometry has to offer.
Proposition 4.2.6. For every Riemannian manifold (M, g), there is a measure
space (Ω, Σ, µ) of smooth functions on M such that length of curves are equal
under Lg and Lµ .
Proof. By the Nash embedding theorem, we may isometrically embed M into some
Rn . A curve in M is then a curve in Rn . We take the same (Ω, Σ, µ) as in example
4.2.5, except that the functions x 7→ x · n + a are restricted to M . (Figure 4.2 shows
how the level curves get induced on the embedded copy of M by the hyperplanes
of Rn .) Then Lg and Lµ are equal up to a scaling factor. Simply rescale µ so that
Lg = Lµ . 

Figure 4.2.
 33

Example 4.2.7. Here is an example which gives a geometry of a different flavor

from the above ones. We consider having  only finitely many functions in Ω. For
concreteness, take M = S n−1 and Ω = x 7→ x · ni | 1 ≤ i ≤ k, n1 , . . . , nk ∈ S n−1
with the counting measure µ imposed. The zero level curves are grand circles and
these grand circles partition the surface of the sphere into pieces (whether an arc
of the grand circle lies in the piece on one side or the other side is determined by
the direction of ni ; we want x · ni ≤ 0 to be one piece and x · ni > 0 to be another
piece). We say that two pieces are adjacent if their boundaries share an arc (of
non-zero length) of one of these grand circles. Under the metric identification of
the pseudometric dµ , we obtain a graph where vertices represent pieces and vertices
are adjacent if and only if the corresponding pieces are adjacent. The path length
between two vertices is equal to k times the distance between any two points in
the corresponding sphere piece under dµ . One might ask: can we characterize the
graphs that are produced by this approach? 
As an idea still in its stage of infancy, it remains to be seen what kinds of
interesting geometries can be obtained on a smooth manifold M by a choice of
(Ω, Σ, µ). We might also examine various types of convergence of our measure µ and
ask if Lµ will respectively converge in some sense. This will give us the possibility
of studying more complicated geometries by approximating with geometries that
we understand. The potential of this framework remains to be seen.
5. Total Curvature Measures
In mathematics you don’t understand things. You just get used to them.
– John von Neumann

In this section, we return to the Euclidean set-up of chapter 3 but leave behind
geometrical notions of size like length and surface area, to give a glimpse of how
integral geometry is a framework that encompasses many other geometrical ideas,
in particular curvature. With regards to curvature, Thomas Banchoff presented
an application of integral geometry that can be found in [Ban67] and [Ban70].
However, this chapter will focus on a different approach that is a direct extension
of the ideas from chapter 3.
The key will again be Steiner’s formula, or rather, a variant of it called the local
Steiner’s formula. We previously saw how the intrinsic volume functions showed up
as the coefficients of the Steiner polynomial, and we were able to interpret directly
some of these coefficients for “sufficiently nice” elements of Kn (e.g. polygons, poly-
hedrons). In turn, this gives us an interpretation of the intrinsic volume functions,
linking their integral geometric definitions with classical terms. We will similarly be
able to interpret the coefficients of the local Steiner polynomial and watch integral
geometric definitions show up as these coefficients.

5.1. Weyl’s Tube Formula. The best way to motivate the approach of this chap-
ter is to realize that the Weyl’s tube formula from differential geometry has a
strikingly similar form to Steiner’s formula. In loose terms, this formula considers
the thickening of hypersurfaces and how the resulting volume is a function of the
amount of thickening. We will present this formula in a slightly different form from
how it is usually stated – a form that is more suggestive of how we should later
extend Steiner’s formula.
For this entire chapter, let n ≥ 2. Recall that for a non-empty compact convex
subset K ∈ Kn , every point x ∈ Rn is associated with a unique closest point in K,
and we write proj : Rn → Rn for this map (the image will be in K). For A ∈ B(Rn )
and  ≥ 0, we define the A-restricted -fattening of K as
K+,A := {x ∈ K+ | proj(x) ∈ A}
For example, K+0,A = K ∩ A and K+,∂K = K+ . The first order of business is to
ensure that we can talk about the volume of these objects.
Proposition 5.1.1. For a fixed K ∈ Kn and  ≥ 0, K+,A is n-Lebesgue measur-
able for every A ∈ B(Rn ). In fact, the function λn (K+,· ) : B(Rn ) → R is a finite
Borel measure (note the · in the subscript of the function λn (K+,· ), which is where
the argument is).
 35

Proof. Observe that the map proj : Rn → Rn is continuous and thus (Lebesgue-
Borel) measurable. Therefore,
K+,A = K+ ∩ proj−1 (A)
is Lebesgue measurable. For the second part of the proposition, observe that the
induced subset σ-algebra of K+ ⊆ Rn is precisely the set of Lebesgue measur-
able subset of Rn contained in K+ (because K+ is itself Lebesgue measurable).
Therefore, it makes sense to talk about the Lebesgue measure induced on K+ .
Now, proj|K+ is (Lebesgue-Borel) measurable and observe that λn (K+,· ) is the
pushforward measure under this map. 
We say that a K ∈ Kn is of class C26=0 if K has non-empty interior and its
boundary ∂K is a C2 -manifold with all of its principal curvatures12 at each point
being non-zero.
Theorem 5.1.2. (Weyl’s tube formula) There are some constants cn,k (which we
will not care about), such that for any K ∈ Kn of class C26=0 and A an open subset13
of Rn ,
n−1
X Z
λn (K+,A \K) = n−k · cn,k Hn,n−1−k dS
k=0 A∩∂K
th
where Hn,k is the k elementrary symmetric function of the n − 1 principal curva-
tures.
Proof. See, for example, [Sch14, (2.63)]. 
Let us make a quick comment about why one would care about the functions
Φn,k (K, ·) : {open subsets of Rn } → R defined by
Z
Φn,k (K, A) := Hn,k dS
A∩∂K
Note that if n = 3 and k = 2, then H3,2 is the usual Gaussian curvature and
Φ3,2 (K, ·) is the usual total curvature function. Thus, we may think of Φn,k (K, ·)
as a generalized sort of total curvature function. Most of these functions do not
have a name, just like how V3,1 on polyhedrons do not have a name. However, these
functions collectively give a more complete description of the curvature behavior of
∂K.
5.2. Local Steiner’s Formula. The polynomial form of Weyl’s tube formula could
not have been more suggestive that one should study, for a general K ∈ Kn , how
λn (K+,A ) behaves as a function of . For a fixed K and a fixed Borel A, the answer
will again be a polynomial of degree ≤ n, but this time the coefficients are functions
of K and A.
Before proceeding, there is an important property about the measures λn (K+,· )
that needs to be recorded as a lemma for future use. The result is intuitive but the
proof will take us too far afield that we shall be content with providing a reference.
If M is a metric space and Σ its Borel σ-algebra, recall that a sequence of its finite
12Principal curvatures refer to the eigenvalues of the Weingarten map (which is the differential
of the Gauss map).
13We chose openness rather than Borel because an open subset intersecting ∂K is an open
submanifold of ∂K for which the surface integral will make sense.
36

signed measures µ1 , µ2 , . . . converges weakly to the finite signed measure µ, denoted

w
µi →µ, if Z Z
f dµi → f dµ
for all bounded, continuous (real-valued) functions f . Recall also the fact from func-
tional analysis is that the weak limit of a sequence of finite signed Borel measures
will be unique. Here is the lemma that we will need.
Lemma 5.2.1. Suppose K, K1 , K2 , · · · ∈ Kn are such that Ki → K under the
Hausdorff metric. Then we have the following weak convergence of measures:
w
λn ((Ki )+,· ) → λn (K+,· )
Proof. See [Sch97, 4.1.1]. 
Let us now proceed to find an expression of λn (K+,A ) in terms of . For the
case of K being a polytope P , a discussion almost identical to that of the original
Steiner’s formula would give us the desired result. Namely, write
a
P+,A = (P+ ∩ proj−1 (relint(F ) ∩ A))
F ∈F

and conclude that

n
!
X X
n−k
λn (P+,A ) =  · λn−k (N (F ) ∩ Bn ) · λk (F ∩ A)
k=0 F ∈Fk

Setting (the notation is chosen to be analogous to that of the original Steiner’s

formula) X
W̃n,k (P, A) := λn−k (N (F ) ∩ Bn ) · λk (F ∩ A)
F ∈Fk

for 0 ≤ k ≤ n − 1 and W̃n,n (P, A) := λn (P ∩ A), we have the following theorem.

Theorem 5.2.2. (Local Steiner’s formula for polytopes) For a polytope P ∈ Pn
and any A ∈ B(Rn ), λn (P+,A ) as a function of  is a ≤ n-degree polynomial. In
fact, with W̃n,0 , . . . , W̃n,n : Pn × B(Rn ) → R defined as above,
n
X
λn (P+,A ) = n−k · W̃n,k (P, A)
k=0

Proof. The proof is given by the discussion above. 

The extension from Pn to Kn is similar in spirit to how it was done for the
original Steiner’s formula, but with a small twist.
Theorem 5.2.3. (Local Steiner’s formula) There are functions W̃n,0 , . . . , W̃n,n :
Kn × B(Rn ) → R such that for any K ∈ Kn , we have
n
X
λn (K+,A ) = n−k · W̃n,k (K, A)
k=0

Proof. As in the proof for theorem 3.4.4, we begin by writing

n
X
W̃n,k (P, A) = ckj · λn (P+j,A )
j=0
 37

where ckj are the entries of the matrix

 −1
1 1 ... 1
 2n 2n−1 ... 1
 
 .. .. .. .. 
 . . . .
nn nn−1 ... 1
Now, for a fixed K ∈ Kn , we already know that λn (K+j,· ) is a finite Borel
measure. Thus, we can define the following (possibly signed) finite Borel measures
n
X
W̃n,k (K, ·) := ckj · λn (K+j,· )
j=0

For a fixed K, choose polytopes P1 , P2 , . . . such that Pi → K (possible by

w
proposition 3.4.2). By lemma 5.2.1, we have λn ((Pi )+,· ) → λn (K+,· ) and thus
w
W̃n,k (Pi , ·) → W̃n,k (K, ·). At the same time, we already have the local Steiner result
for polytopes (theorem 5.2.2), so
n n
w
X X
λn ((Pi )+,· ) = n−k · W̃n,k (Pi , ·) → n−k · W̃n,k (K, ·)
k=0 k=0
By uniqueness of weak limit, we conclude that
n
X
λn (K+,· ) = n−k · W̃n,k (K, ·)
k=0
That is,
n
X
λn (K+,A ) = n−k · W̃n,k (K, A)
k=0
for all A ∈ B(Rn ). 
By a direct comparison of Weyl’s tube formula and local Steiner’s formula, we
see that if in particular K is of class C26=0 and A is open, then W̃n,k (K, A) is simply
R
a rescaled version of Φn,n−1−k (K, A) = A∩∂K Hn,n−1−k dS, where the constant
of proportionality is independent of K and A. Thus, we can directly interpret the
coefficients of the local Steiner polynomial as (generalized) total curvatures.
5.3. V -measures and Integral Geometry. We will now use integral geometry
to define a class of Borel measures called V -measures, so named because they are a
simple modification of the intrinsic volume functions of chapter 3. We will introduce
a second argument, writing Vn,k (K, A) where A ∈ B(Rn ). (This is slight overuse of
notation but we will soon see that Vn,k (K) = Vn,k (K, Rn ).) For a fixed K, it will
turn out that Vn,k (K, ·) is a Borel measure.
Compared to the order of presentation in chapter 3, the appearance of integral
geometry has been delayed till the end (i.e. now) because it would not have been
easy to motivate why anyone would be interested in looking at these V -measures.
The definitions will appear less arbitrary if one keeps the local Steiner’s formula in
mind – in the end, these V -measures appear as the coefficients in the local Steiner
polynomial.
From the onset, let us be clear that to work out some technical details of this
section requires a deep journey into the study of compact convex bodies, and can
38

be enough material for another thesis. Our main goal is to give a glimpse of how
integral geometry can be connected to curvature. To focus on the key integral
geometric ideas, we would simply mention the details that need to be checked and
refer interested readers to [Sch14], an extensive monograph on compact convex
bodies.
Recall that in chapter 3, the function Vn,0 is essentially an indicator of whether
the set is empty or not. Crofton’s formula (proposition 3.2.4) then define the other
functions Vn,1 , . . . , Vn,n in terms of some expectation involving Vn,0 . Motivated by
the study of the local Steiner’s formula, we now define the V -measure Vn,0 to care
only about points of ∂K ∩ A as follows.
Definition 5.3.1. Let K ∈ Kn , A ∈ B(Rn ). Let N (K, A) ⊆ Rn be the set
consisting of the zero vector and all (not necessarily unit) outer normal vectors of
K at points of ∂K ∩ A. Define
λn (N (K, A) ∩ B n )
Vn,0 (K, A) :=
λn (B n )
where λn (B n ) is just a normalization factor. For this definition to make sense, one
should verify that N (K, A) ∩ B n is always n-Lebesgue measurable. One can also
check that Vn,0 (K, ·) satisfies the axioms of a measure.
Note that for non-empty K ∈ Kn , the set of all its outer unit normal vectors,
when varied across its entire boundary, is S n−1 . Thus, Vn,0 (K, Rn ) = 1 = Vn,0 (K)
for non-empty K. Intuitively, Vn,0 (K, A) is now the “fraction” of ∂K that also lies
in A, judged by looking at the set of outer unit normal vectors of ∂K ∩ A as a
fraction of the set of outer unit normal vectors of ∂K. For some motivation on why
such a quantity might be related to curvature, think of how a patch with high total
curvature will have unit normal vectors that sweep through a larger proportion of
S n−1 .

Figure 5.1.
 39

Example 5.3.2. An interesting example to study would be where K is a polytope

P . Recall that we also used the notation N (F ) for the normal cone of F . In this
case, note that the vertices of P are the only ones that “matter” (see figure 5.1).
More specifically,
X λn (N (v) ∩ B n )
Vn,0 (P, A) =
λn (B n )
v∈F0 ,v∈A

This is true even if P is a “lower dimensional” polytope. One can think of the
contributions to Vn,0 (P, ·) as being concentrated at the vertices of P , with the
contribution being effected if and only if the vertex is in A. Later when we show the
link between Vn,0 and total curvature, one can think of all the “Gaussian curvature”
of P being solely concentrated at the vertices. 
Here is the integral geometric definition of the other V -measures.
Definition 5.3.3. (Local Crofton formula) For 1 ≤ k ≤ n, define
Z
Vn,k (K, A) := Vn,0 (K ∩ P, A) αn,n−k (dP )
Aff(n,n−k)

As usual, one has to check that Vn,0 (K ∩ ·, A) is a measurable function. One can
also check that Vn,k (K, ·) satisfies the axioms of a measure.
Note that
Z
Vn,k (K, Rn ) = Vn,0 (K ∩ P ) αn,n−k (dP ) = Vn,k (K)
Aff(n,n−k)

Intuitively, Vn,k (K) previously counts each intersection with a plane “fully”, but
Vn,k (K, A) counts only a “fraction” of every intersection, by looking at the “frac-
tion” of ∂(K ∩ P ) that also lies in A, via the use of Vn,0 . Note that Vn,n (P, A) =
λn (K ∩ A).
Example 5.3.4. For the case of K being a polytope P , it turns out that
X λn−k (N (F ) ∩ B n )
Vn,k (P, A) = cn,k · λk (F ∩ A)
λn−k (B n−k )
F ∈Fk (P )

where cn,k are constants independent of P and A. Compared to example 5.3.2,

this example is more messy to show rigorously, but figure 5.2 illustrates (with
n = 3, k = 1) why the above form is intuitively believable. Intuitively, each
intersection of P with a (n−k)-plane will generally be (n−k)-dimensional polytope,
and its Vn,0 output is determined by its vertices (example 5.3.2), which are actually
points on the k-face of the original polytope P .
Later when we show the link between V3,1 and integral of mean curvature, one
can think of all the “mean curvature” of P as being solely concentrated at the edges
of the cube. 
Finally, this proposition will provide the link between V -measures and (general-
ized) total curvature.
Proposition 5.3.5. The V -measures Vn,k are respectively equal to the coefficients
of the local Steiner polynomial W̃n,k , up to a scaling factor.
40

plane

Figure 5.2.

Proof. (Sketch) Suppose K, K1 , K2 , · · · ∈ Kn are such that Ki → K under the

Hausdorff metric. In the course of proving theorem 5.2.3, we saw that lemma 5.2.1
w
implied W̃n,k (Ki , ·) → W̃n,k (K, ·). It can be shown that a similar weak convergence
w
property holds for the V -measures, namely Vn,k (Ki , ·) → Vn,k (K, ·). If we further
believe the formula for polytopes discussed in example 5.3.4, we see that W̃n,k (P, ·)
and Vn,k (P, ·) agree whenever P is a polytope. For a fixed K, choose polytopes
P1 , P2 , . . . such that Pi → K and by uniqueness of weak limits, we conclude that
W̃n,k (K, ·) = Vn,k (K, ·). 
Just as how the intrinsic volume functions can be thought of as notions of size
applicable to compact convex sets, the V -measures can be thought of as notions
of total curvature applicable to compact convex sets and whose measure domain
applies to all Borel subsets of Rn , not just open ones. Therefore, we can talk about
total curvature of patches of ∂K that look nothing like a manifold.
6. Recent Developments
The worst thing you can do is to completely solve a problem.
– Daniel Kleitman

To conclude, we list a number of recent developments in integral geometry. Many

of these extensions have been motivated by applications in stochastic geometry
[SW08]. This list is by no means exhaustive – we only cover extensions that are
more directly related to the content discussed in this thesis. A more thorough
survey that also covers other aspects of integral geometry can be found in [HS02].
More general sets. While convex sets (or even polyconvex sets) constitute
a wide collection of objects, there have been further attempts to establish some
form of Crofton’s formula or curvature measures on collections of sets with more
pathological behaviors. For example, Fu [Fu94] examined “subanalytic sets” while
Brocker and Kuppe [BK00] further advanced this subject by considering “compact
tame Whitney-stratified sets” (this collection includes semi-algebraic sets, suban-
alytic sets, and sets belonging to an σ-minimal system). These sets show up in
model theory and computer science.
Non-Euclidean Spaces. Starting from a given non-Euclidean space X, one
can focus on a collection of sets K ⊆ 2X and ask if we can similarly14 characterize
valuations on K in integral geometric terms. Little is known in the general context
but Klain and Rota [KR97, chapters 3 and 11] provides some answers for the case
of X being a sphere and X being a finite set.
Crofton formula for functions. A function f : Ω → R, where Ω ⊆ Rn is open
and convex, is a convex function if and only if its epigraph is convex. This suggests
the possibility of studying such a convex function by the techniques we explored.
Indeed, for  ≥ 0 and A is Borel set contained in Ω, one may define the -fattening
of A by means of f as
{x + v | x ∈ A, v ∈ ∂f (x)}
where ∂f (x) denotes the “subdifferential” of f at x. It turns out that the fattening
is a Lebesgue measurable subset of Rn , and its Lebesgue measure as a function of
 is a polynomial of degree ≤ n. By considering the special case where f happens
to be of class C2 , these coefficients can be interpreted in terms of some expression
involving the eigenvalues of Hess(f ) at various points (the coefficients are thus
called Hessian measures). One can make various integral geometric definitions of
measures that turn out to be Hessian measures. A concrete discussion can be
found in [CH05]. For comparison, a study of Hessian measures without the integral
geometric approach can be found in [TW97], [TW99] and [TW02].
Translative integral geometry. In some applications of stochastic geometry,
an assumption that space is isotopic is superfluous or invalid. Thus, rather than
14We characterized all continuous invariant valuations when X = Rn and K = Kn as linear
combinations of intrinsic volume functions.
42

considering invariance under E (n), one only considers invariance under the transla-
tion group T (n). The study of expectations under this setting leads to translative
integral geometry. This subject is treated in [AF14, 1.1].
 43

References
[AF14] S. Alesker and J.H.G. Fu. Integral Geometry and Valuations. Springer, 2014.
[Ban67] T. Banchoff. Critical points and curvature for embedded polyhedra. J. Differential Ge-
ometry, 1:245–256, 1967.
[Ban70] T. Banchoff. Critical points and curvature for embedded polyhedral surfaces. Amer.
Math. Monthly, 77:475–485, 1970.
[Bee74] G.A. Beer. The Hausdorff metric and convergence in measure. Michigan Math. J.,
21(1):63–64, 1974.
[BK00] L. Bröcker and M. Kuppe. Integral geometry of tame sets. Geom. Dedicata, 82:285–328,
2000.
[CH05] A. Colesanti and D. Hug. Hessian measures of convex functions and area measures. J.
London Math. Soc., 71:221–235, 2005.
[DL94] R. De-Lin. Topics in Integral Geometry. World Scientific Pub Co Inc, 1994.
[DO11] S.L. Devadoss and J. O’Rourke. Discrete and Computational Geometry. Princeton Uni-
versity Press, 2011.
[Fol99] G.B. Folland. Real Analysis: Modern Techniques and their Applications. John Wiley &
Sons, Inc., second edition, 1999.
[Fu94] J.H.G. Fu. Curvature measures of subanalytic sets. Amer. J. Math, 116:819–880, 1994.
[HS02] D. Hug and R. Schneider. Kinematic and crofton formulae of integral geometry: recent
variants and extensions. Homenatge al professor Llus Santal i. Sors, C. Barcel i Vidal
(ed.):51–80, 2002.
[Kla95] D.A. Klain. A short proof of Hadwiger’s characterization theorem. Mathematika,
42(2):329–339, 1995.
[Kni] O. Knill. Integral geometric kolmogorov quotients. In preparation.
[Kni12] O. Knill. The Theorems of Green-Stokes,Gauss-Bonnet and Poincare-Hopf in Graph
Theory. http://arxiv.org/abs/1201.6049, 2012.
[Kni14] O. Knill. Curvature from graph colorings. http://arxiv.org/abs/1410.1217, 2014.
[KQ03] P. Kuchment and E.T. Quinto. Some problems of integral geometry arising in tomogra-
phy. In L. Ehrenpreis, editor, The Universality of the Radon Transform. Oxford Univer-
sity Press, London, 2003.
[KR97] D.A. Klain and G-C. Rota. Introduction to Geometric Probability. Lezioni Lincee. Ac-
cademia nazionale dei lincei, 1997.
[Kuh06] W. Kuhler. Differential Geometry: Curve-Surfaces-Manifolds. American Mathematical
Society, second edition, 2006.
[Lee12] J.M. Lee. Introduction to Smooth Manifolds. Graduate Texts in Mathematics. Springer,
second edition, 2012.
[Lei11] T. Leinster. Hadwiger’s theorem, part 1. https://golem.ph.utexas.edu/category/2011/
06/hadwigers theorem part 1.html, 2011.
[Mec98] K.R. Mecke. Integral geometry in statistical physics. International Journal of Modern
Physics B, 12(9):861–899, 1998.
[Mil50] J. Milnor. On the total curvature of knots. Annals of Mathematics, 52:248–257, 1950.
[Ram69] J.F. Ramaley. Buffon’s noodle problem. Amer. Math. Monthly, 76(8):916–918, 1969.
[San53] L.A. Santalo. Introduction to Integral Geometry. Hermann and Editeurs, Paris, 1953.
[San04] L.A. Santalo. Integral Geometry and Geometric Probability. Cambridge University Press,
second edition, 2004.
[Sch07] R. Schneider. Integral geometric tools for stochastic geometry. Lecture Notes in Mathe-
matics, 1892:119–184, 2007.
[Sch14] R. Schneider. Convex Bodies: the Brunn-Minkowski Theory. Cambridge University
Press, second expanded edition, 2014.
[SW08] R. Schneider and W. Weil. Stochastic and Integral Geometry. Springer, 2008.
[Tis84] P.E. Tissier. Bertrand’s paradox. The Mathematical Gazette, 68(443):15–19, 1984.
[Tor14] S. Torniers. Haar measures. https://people.math.ethz.ch/ torniers/download/2014/
haar measures.pdf, 2014.
44

[TW97] N.S. Trudinger and X.J. Wang. Hessian measures I. Topol. Methods Nonlinear Anal.,
10:225–239, 1997.
[TW99] N.S. Trudinger and X.J. Wang. Hessian measures II. Ann. of Math., 150:579–604, 1999.
[TW02] N.S. Trudinger and X.J. Wang. Hessian measures III. Journal of Functional Analysis,
193(1):1–23, 2002.
[Vin08] R. Vinroot. Topological groups. http://www.math.wm.edu/∼vinroot/PadicGroups/ top-
groups.pdf, 2008.
[Wey39] H. Weyl. On the volume of tubes. American Journal of MAthematics, 61(2):461–472,
1939.