Euler Calculus and Applications

Kevin Gong and Helen Jin

March 14, 2019

1 Brief Review of Euler Characteristic

Recall that the Euler characteristic is a generalization of counting. Given a finite set X, the Euler charac-

teristic is its cardinality χ(X) = |X|.

The first way one thinks about Euler characteristic is as follows: if one connects two points of X together

by means of an edge (in a cellular/simplicial structure), the resulting space has one fewer component and the

Euler characteristic is decremented by one. Continuing inductively, the Euler characteristic counts vertices

with weight +1 and edges with weight -1.

This intuition of counting connected components works at first; however, the addition of an edge produc-

ing a cycle does not change the count of connected components. To fill in such a cycle with a 2-cell would

return to the setting of counting connected components again, suggesting 2-cells be weighted with +1. This

inspires the combinatorial definition of the Euler characteristic.

Definition. Given a space X and a partition thereof into a finite number of open cells X = tα σα , where

each k-cell σα is homeomorphic to Rk , the Euler characteristic of X is defined as

X
χ(X) := (−1)dim σα
α

For an appropriate class of ”tame” spaces, this quantity is well-defined and independent of the cellular

decomposition of X. We will spend time in the next section to understand what ”tame” really means.

Note that Euler characteristic is a homeomorphism invariant, but as defined, is not a homotopy invariant

for non-compact spaces, as e.g., it distinguishes χ((0, 1)) = −1 and χ([0, 1]) = 1. Among compact finite cell

complexes though, it is a homotopy invariant.

1
2 O-minimal structure, Tame Topology and Definable Sets

Definition. A o-minimal structure O = {On } (over R) is a sequence of Boolean algebras On of subsets

of Rn (families of sets closed under the operations of intersection and complement) which satisfies certain

axioms:

1. O is closed under cartesian products;

2. O is closed under axis-aligned projections Rn → Rn−1 ;

3. O contains all algebraic sets (zero-sets of polynomials);

4. O1 consists of all finite unions of points and open intervals.

Elements of O are called tame, or more properly, definable sets. Canonical examples of o-minimal struc-

tures are semialgebraic sets (expressible in terms of a finite set of polynomial inequalities) and subanalytic

sets (defined in terms of images of analytic mappings).

An example of a set that is not tame is sin(1/x) for 0 < x < 1. Though this is homeomorphic to an

interval, the closure of this set in R2 is not homeomorphic to a closed interval.

Figure 1: graph of sin(1/x), 0 < x < 1

Definitions.

• A tame (or definable) function is a function between tame spaces whose graph (in the product of

domain and range) is a tame set.

• A definable homeomorphism is a tame bijection between tame sets.

Note: these do not necessarily have to be continuous.

2
Triangulation Theorem. Any tame set is definably homeomorphic to a finite disjoint union of open

standard simplices. The intersection of the closures of any two of the simplices in this definable triangulation

is either empty, or the closure of another open simplex in the triangulation.

Figure 2: an example of the Triangulation Theorem

This result implies that tame sets always have a well-defined Euler characteristic and a well-defined

dimension (the max of the dimensions of the simplices in a triangulation). These two quantities are not just

topological invariants with respect to definable homeomorphism, but they are complete invariants.

Invariance Theorem. Two tame sets in an o-minimal structure are definably homeomorphic if and only

if they have the same dimension and Euler characteristic.

This result reinforces the idea of a definable homeomorphism as a scissors equivalence (i.e. one is

permitted to cut and rearrange a space with abandon).

3
3 Euler Calculus

It is possible to build a topological calculus based on Euler characteristic. We start off with a lemma that

is essential to the integral of this calculus.

Lemma. For definable sets A and B,

χ(A ∪ B) = χ(A) + χ(B) − χ(A ∩ B).

This can be proved via triangulation, induction, and toil, or by other advanced topological tools but will

be skipped for now.

Figure 3: represents definable sets A and B and their relationship

Definition. We can construct a measure dχ over definable sets A ⊂ X via:

Z
1A dχ := χ(A)
X

Definition. Measurable functions in this integration theory are integer-valued and constructible, meaning

that for h : X → Z, all level sets h−1 (n) ⊂ X are tame. Denote CF(X) the set of bounded compactly

supported constructible functions on X.

R
The Euler integral is defined to be the homomorphism X
: CF(X) → Z given by:

Z ∞
X
hdχ = sχ({h = s})
X s=−∞

4
Proposition. For any h ∈ CF(X),

Z ∞
X
hdχ = χ({h > s}) − χ({h < −s})
X s=0

Proof. Rewrite h as:

∞ ∞ −∞ ∞
s1{h=s} = s(1{h≥s} − 1{h>s} ) + s(1{h≤s} − 1{h<s} ) = 1{h>s} − 1{h<−s}
X X X X
h=
s=−∞ s=0 s=0 s=0

where the last equality comes from telescoping sums.

(One can see that this proposition also utilizes a manifestation of a discrete fundamental theorem of integral

calculus.)

1
P
Definition. Alternatively, one can write h ∈ CF(X) as h = α cα σα , where cα ∈ Z and {σα } is a

decomposition of X into a disjoint union of open cells, yielding:

Z X X
h dχ = cα χ(σα ) = cα (−1)dimσα
X α α

That this sum is invariant under the decomposition into definable cells is a consequence of corresponding

properties of the Euler characteristic.

Figure 4: A simple example of an integrand h ∈ CF (R2 ) whose Euler integral equals six: it is the sum of

six characteristic functions over contractible discs in the plane.

5
4 Target enumeration

A simple application of Euler calculus is target enumeration, used mostly in engineering disciplines.

Consider a finite collection of targets, represented as discrete points in a space W . Assume a field of sensors,

each of which observes some subset of W and counts the number of targets therein. The sensors will be

assumed to be distributed so densely as to be approximated by a topological space X.

There are many modes of sensing: infrared, acoustic, optical, magnetometric, etc. We will focus on a

topological approach.

Definitions.

• In a particular system of sensors in X and targets in W , let the sensing relation be the relation

S ⊂ W × X where (w, x) ∈ S iff a sensor at x ∈ X detects a target at w ∈ W .

• The vertical fibers (inverse image of the projection of S to W ) are target supports, sets of sensors

which detect a given target in W .

• The horizontal fibers (inverse image of the projection of S to X) are sensor supports, the target

locations observable by a given sensor in X.

Figure 5: A graph of the sensing relation S ⊂ W × X and a general view of the sensors and targets in the
region

Assume sensors are additive but anonymizing: each sensor at x ∈ X counts the number of targets and

returns a local count h(x), but the identities of the targets sensed are unknown. This counting function

h : X → Z is, under the usual tameness assumptions, constructible.

6
A natural problem in this context is to aggregate the redundant anonymous target counts: given h and

some minimal information about the sensing relation S, determine the total number of targets in W . This

is in essence a problem of computing a global section (number) from a collection of local sections (target

counts).

Proposition. If h : X → N is a counting function for target supports Uα of uniform Euler characteristic

χ(Uα ) = N 6= 0 for all α, then

Z
1
#α = h dχ
N X

Proof.
Z Z XZ
1Uα ) dχ = ( 1Uα dχ) =
X X
h dχ = ( χ(Uα ) = N #α
X X α α X α

Notice that for contractible supports, the target count is, simply, the Euler integral of the function. This

solves a problem in the aggregation of redundant data, since many nearby sensors with the same reading are

detecting the same targets; in the absence of target identification (which is an expensive signal processing

task), it is nontrivial to aggregate the redundacy.

Note that the restriction N 6= 0 is not trivial. For example, if h ∈ CF(R2 ) is a finite sum of charac-
R
teristic functions over annuli, it is not merely inconvenient that R2 h dχ = 0, it is actually a fundamental

obstruction to disambiguating sets. Some sums of annuli may be expressed as a union of different numbers

of embedded annuli.

Figure 6: An example for which the integral with respect to Euler characteristic vanishes. Note that the

integrand can be represented as a sum of indicator functions of embedded annuli, the number of which is

indeterminate.

7
5 Fubini

We will discuss some of the properties and transformations of Euler calculus. First, similarly to in multi-

variable calculus, we have another formulation of the Fubini Theorem. Roughly, the theorem says to first

integrate over the fibers (level-sets) and then integrate over the base space.

Lemma For X, Y definable sets, χ(X × Y ) = χ(X)χ(Y )

Proof First, X ×Y is also definable and has a definable cell structure coming from the product of cells in X

and Y. Then let σ ⊂ X and τ ⊂ Y both be cells. Then dim(σ×τ ) = dim(σ)+dim(τ ), so χ(σ×τ ) = χ(σ)χ(τ ).

Using additivity, we get that χ(X × Y ) = χ(X)χ(Y )

Definitions First, let A ⊂ Rm be definable and X ⊂ Rm × Rn be a definable family of subsets of Rn .

Then, say X is definably trivial over A if there exists a set F definable and definable homeomorphism

h : A × F → X|A = X ∩ (A × Rn ) such that the following diagram commutes:

Then, say that the trivialization h is compatible with a definable subset Y ⊂ X if there is a definable

subset G ⊂ F such that h(A × G) = Y |A

Hardt Theorem The Hardt theorem is typically stated using semialgebraic sets, but it is possible to

extend the Hardt Theorem to o-minimal structures. Let X ⊂ Rm × Rn be a definable family and Y1 , ..., Yl

be definable subsets of X. Then, there exists a finite partition of Rm into definable subsets C1 , ..., Ck such

that X is definably trivial over each Ci and trivializations over each Ci are compatible with Y1 , ..., Yl .

Vague Sketch of Proof Using triangulation, find a trivialisation h. Then use compactness.

Fubini Theorem: Let F : X → Y be a definable mapping. Then ∀h ∈ CF (X)

Z Z Z
hdχ = h(x)dχ(x)dχ(y)
χ Y F −1 (y)

8
Proof If X = U × Y and F : U × Y → Y be the projection, then the result simply follows from the

Lemma. Otherwise, using the Hardt Theorem, Y has a definable partition in definable Yα such that F −1 (Yα )

is homeomorphic to Uα × Yα , where Uα is definable and F : Uα × Yα → Yα is the projection. Then, additivity

of the integral completes the proof.

6 Integral Transforms

6.1 Convolution and Minkowski Sums

Definition On a real vector space V and f, g ∈ CF (V ) a convolution with respect to the Euler charac-

teristic is:
Z
(f ∗ g) = f (t)g(x − t)dχ(t)
V
R R R
Thus, we can see that f ∗ gdχ = f dχ gdχ, using Fubini’s Theorem.

The convolution has a close relationship with the Minkowski Sum, which for convex sets is defined as A, B,

1A ∗ 1B = 1A+B , where A + B are all vectors expressible as a sum of vectors of A plus vectors of B. If A and

B are not convex, then the convolution of indicator functions can be larger than 1, where the intersections

of translations of A and B are disconnected. However, the Euler-convolution is always invertible, unlike the

non-convex Minkowski sum.

6.2 Radon Transformations

Definition For a locally closed, definable relation S ⊂ W × X. The Radon Transform with kernel S is

RS : CF (W ) → CF (X) given by lifting h ∈ CF (W ) from W to W × X, filtering with the kernel 1S , then

integrating along the projection to X as follows:

Z
(RS h)(x) := h(w)1S (x, w)dχ(w)
W

It is easier to see/understand the Radon transformation from the following diagram.

9
Thus, an easy example is that the counting function which the sensor field on X returns is the Radon

Transform RS 1T .

Schapira inversion formula Assume that S ⊂ W × X and S 0 ⊂ X × W have fibers Sw and S 0 w ,

0 0 0
respectively, such that χ(Sw ∩ Sw ) = µ for all w ∈ W and χ(Sw ∩ Sw 0 ) = λ for all w 6= w ∈ W . Then

∀h ∈ CF (W ),
Z
(RS 0 ◦ RS )h = (µ − λ)h + λ h ∗ 1W
W

Topological tomography Let W = R3 and compact T ⊂ W obtained by slicing through R3 through flat

hyperplanes of R3 . Then, record the Euler characteristic of each T. Then, it is possible to reconstruct an

accurate Euler characteristic of the entire space, even with noisy readings. Formally let X be the sensor space

AGr23 , which is the affine Grassmannian manifold of 2-planes in R3 ). Then, we have a constructible function

h on X, which is equal to the Radon transform of 1T (RS 1T ). Then, we can use the Schapira inversion

= P2 , (SW ∩ SW 0 ) ∼
formula, with the following facts: SW ∼ = P1 , therefore, µ = χ(P2 ) = 1 and λ = χ(P1 ) = 0.
Thus, the inverse Radon transform gives 1T exactly.

10
7 Intrinsic Volumes

Definition The kth intrinsic volume, µk is characterized uniquely by the following: for all A, B ⊂ Rn

definable,

• Additivity: µk (A ∪ B) = µk (A) + µk (B) − µk (A ∩ B)

• Euclidean invariance: µk is invariant under rigid motions of Rn

• Homogeneity under scaling: µk (cA) = ck ∗ A, for all c ≥ 0

• Normalization: µk (B1n [0]) = 1, where B1n [0] is the closed unit ball in Rn

There are many different formulations for µn , which generalize the Euler characteristic (µ0 = dχ and the

n-dimensional volume µn = dvoln . For instance, in R3 , µ0 = dχ, µ1 is the “length”, µ2 is the surface area,

and µ3 is the volume. Thus, for a unit sphere, µ0 = 2, µ1 = 2π, µ2 = 4π, µ3 = 43 π. For A definable, one

way of formally defining the intrinsic volume µk (A) is using the Euler characteristic of all slices of A along

affine codimension-k planes:

Z Z Z
µk (A) = χ(A ∩ P )dλ(P ) = 1A dχdλ(P )
AGk
n−k AGk
n−k P

where dλ is a (appropriate) measure on AGkn−k (which is derived from the Haar measure on the Grassmanian

Gkn−k and the Lebesgue measure on the orthogonal Rk ). Then,

Hadwiger Theorem The space of all compact-continuous Euclidean-invariant valuations on Rn is a vector

space of dimension n + 1 with basis {µk }nk=0 .

Instead of a proof, I will give some examples.

• R0 is a singleton, so we have two convex subsets φ, R0 . Since the valuation of the empty set must be

0, then we can value R0 to be any real number, so V al0 ∼

= R =⇒ dim(V al0 ) = 1.

• R1 has the Euler characteristic and Lebesgue measure, which are linearly independent since they are

homogeneous of degree 0 and 1, respectively. Thus dim(V al1 ) ≥ 2. But, convex subsets of R are

compact intervals (and the empty set), so any continuous invariant valuation is determined by its

valuation on {0} and [0, 1]. Thus, dim(V al1 ) ≤ 2 =⇒ dim(V al1 ) = 2.

• R2 has the Euler characteristic, perimeter and Lebesgue measure (area). Again these are linearly

independent by homogeneity so dim(V al2 ) ≥ 3. Then convince yourself that any continuous invariant

valuation on R2 is a linear combination of the above, so that dim(V al2 ) = 3.

11
Finally, a note from before, in R3 , the 1-dimensional measure is actually the called the mean width and

is constructed as follows by taking a random line l going through the origin, projecting the space X onto l

using πl X and measuring the distance along that line.

8 Extension to R

Currently, the Euler integral has a range defined for just the integers, but it is possible to extend this measure

to the reals, “For purposes of constructing an honest numerical analysis for the constructible setting of Euler

integration”. First let Def (X, Y ) be the set of definable functions f : X → Y whose graphs in X × Y

are tame and let Def k (X, Y ) be the maps which are C k smooth. Then, we can define Euler integrals over

Def (X) = Def (X, R) using a Riemann-sum definition. Given h ∈ Def (X), let

Z Z Z Z
1 1
h bdχc = lim bnhc dχ and h ddχe = lim dnhe χ.
X n→∞ n x X n→∞ n x

These integrals are well-defined, given an extension of the Triangle Theorem to Def (X). For each open

k-simplex σ, σ h bdχc = (−1)k inf σ h and σ h ddχe = (−1)k supσ h. But, furthermore, the two different
R R

formulations are not equal in general.

However, these integrals are not linear, in general since the floor and ceiling functions are not linear. This

can be seen in an explicit example:

Z Z Z
1= 1 bdχc =
6 x bdχc + 1 − x bdχc = 1 + 1 = 2
[0,1] [0,1] [0,1]

Fubini Theorem Furthermore, the Fubini Theorem also fails in general because of this fact, but the

Fubini theorem does hold when the map respects the fibers, i.e. For h ∈ Def (X), let F : X → Y be
R R
definable and h-preserving (i.e. h is constant on the fibers of F). Then Y F∗ h = X h, with respect to bdχc

or ddχe. Proof: Similar as before

12
Finally, the two different integral formulations are very closely related by the following relationship, which

is simplied derived using triangulation:

Z Z
h ddχe = − −h bdχc

9 References

• https://golem.ph.utexas.edu/category/2011/06/hadwigers theorem part 1.html

• https://www.math.upenn.edu/ ghrist/preprints/eulertome.pdf

• https://www.math.upenn.edu/ ghrist/EAT/EATchapter3.pdf

13