Fundamental

SOME FUNDAMENTAL THEOREMS IN MATHEMATICS
OLIVER KNILL
Abstract. An expository hitchhikers guide to some theorems in mathematics.

Criteria for the current list of 172 theorems are whether the result can be formulated elegantly,
whether it is beautiful or useful and whether it could serve as a guide [6] without leading to
panic. The order is not a ranking but ordered along a time-line when things were written
down. Since [348] stated “a mathematical theorem only becomes beautiful if presented as a
crown jewel within a context” we try sometimes to give some context. Of course, any such list
of theorems is a matter of personal preferences, taste and limitations. The number of theo-
rems is arbitrary, the initial obvious goal was 42 but that number got eventually surpassed as
it is hard to stop, once started. As a compensation, there are 42 “tweetable” theorems with
included proofs. More comments on the choice of the theorems is included in an epilogue.
For literature on general mathematics, see [132, 128, 23, 156, 167, 391, 265, 95], for history
[147, 396, 240, 49, 37, 139, 242, 234, 434, 78, 390, 54, 170, 218], for popular, beautiful or elegant
things [9, 330, 135, 123, 14, 422, 423, 35, 136, 129, 162, 283, 388, 195, 135, 2, 86, 102, 87, 315]. For
comprehensive overviews in large parts of mathematics, [50, 114, 115, 40, 371] or predictions on
developments [38]. For reflections about mathematics in general [101, 288, 36, 197, 281, 70, 352].
Encyclopedic source examples are [127, 440, 420, 71, 131, 106, 150, 130, 76, 399].
This is still a live document and will be extended. Thanks to Jordan Stoyanov for some valuable
comments and corrections.
1. Arithmetic
Let N = {0, 1, 2, 3, . . . } be the set of natural numbers. A number p ∈ N, p > 1 is prime
if p has no factors different from 1 and p. With a prime factorization n = p1 . . . pn , we
understand the prime factors pj of n to be ordered as pi ≤ pi+1 . The fundamental theorem
of arithmetic is
Theorem: Every n ∈ N, n > 1 has a unique prime factorization.
Euclid anticipated the result. Carl Friedrich Gauss gave in 1798 the first proof in his monograph
“Disquisitiones Arithmeticae”. Within abstract algebra, the result is the statement that the
ring of integers Z is a unique factorization domain. For a literature source, see [227]. For
more general number theory literature, see [204, 79].
2. Geometry
√
Given an inner product space (V, ·) with dot product v · w leading to length |v| = v.v,
three non-zero vectors v, w, v − w define a right angle triangle if v and w are perpendicular
Date: 7/22/2018, last update 6/19/2019.

1
FUNDAMENTAL THEOREMS
meaning that v · w = 0. If a = |v|, b = |w|, c = |v − w| are the lengths of the three vectors, then
the Pythagoras theorem is
Theorem: a2 + b2 = c2 .
Anticipated by Babylonians Mathematicians in examples, it appeared independently also in
Chinese mathematics [397] and was proven first by Pythagoras. It is used in many parts of
mathematics like in the Perseval equality of Fourier theory. See [334, 285, 234].
3. Calculus
Let f be a function of one variables which is continuously differentiable, meaning that
the limit g(x) = limh→0 [f (x + h) − f (x)]/h exists at every point x and defines a continuous
Rb
function g. For any such function f , we can form the integral a f (t) dt and the derivative
d/dxf (x) = f 0 (x).
Rb Rx
Theorem: a
f 0 (x)dx = f (b) − f (a), d
dx 0
f (t)dt = f (x)
Newton and Leibniz discovered the result independently, Gregory wrote down the first proof in
his “Geometriae Pars Universalis” of 1668. The result generalizes to higher dimensions in the
form of the Green-Stokes-Gauss-Ostogradski theorem. For history, see [233]. [134] tells
the “tongue in the cheek” proof: as the derivative is a limit of quotient of differences, the
anti-derivative must be a limit of sums of products. For history, see [133]
4. Algebra
A polynomial is a complex valued function of the form f (x) = a0 + a1 x + · · · + an xn , where
the entries ak are in the complex plane C. The space of all polynomials is denoted C[x]. The
largest non-negative integer n for which an 6= 0 is called the degree of the polynomial. Degree
1 polynomials are linear, degree 2 polynomials are called quadratic etc. The fundamental
theorem of algebra is
Theorem: Every f ∈ C[x] of degree n can be factored into n linear factors.
This result was anticipated in the 17th century. The first writer to assert that any n’th degree
polynomial has a root is Peter Roth in 1600 [332], proven first by Carl Friedrich Gauss and
finalized in 1920 by Alexander Ostrowski who fixed a topological mistake in Gauss proof. The
theorem assures that the field of complex numbers C is algebraically closed. For history and
many proofs see [146].
5. Probability
Given a sequence Xk of independent random variables on a probability space (Ω, A, P)
which all have the same cumulative distribution functions FX (t) = P[X ≤R t]. The nor-
malized random variable X = is (X − E[X])/σ[X], where E[X] is the mean Ω X(ω)dP (ω)
and σ[X] = E[(X − E[X])2 ]1/2 is the standard deviation. A sequence of random variables
Zn → Z converges in distribution to Z if FZn (t) → FZ (t) for all t as n → ∞. If Z is a
Gaussian random variable with zero mean E[Z] = 0 and standard deviation σ[Z] = 1, the
central limit theorem is:
2
OLIVER KNILL
Theorem: (X1 + X2 + · · · + Xn ) → Z in distribution.
Proven in a special case by Abraham De-Moivre for discrete random variables and then by
Constantin Carathéodory and Paul Lévy, the√theorem explains the importance and ubiquity
2
of the Gaussian density function e−x /2 / 2π defining the normal distribution. The
Gaussian distribution was first considered by Abraham de Moivre from 1738. See [395, 248].
6. Dynamics
Assume X is a random variable on a probability space R (Ω, A, P) for which |X| has finite
mean E[|X|]. This means X : Ω → R is measurable and Ω |X(x)|dP(x) is finite. Let T be an
ergodic, measure-preserving transformation from Ω to Ω. Measure preserving means that
P [T −1 (A)] = P [A] for all measurable sets A ∈ A. Ergodic means that that T (A) = A
implies P[A] = 0 or P[A] = 1 for all A ∈ A. The ergodic theorem states, that for an ergodic
transformation T on has:
Theorem: [X(x) + X(T x) + · · · + X(T n−1 (x))]/n → E[X] for almost all x.
This theorem from 1931 is due to George Birkhoff and called Birkhoff ’s pointwise ergodic
theorem. It assures that “time averages” are equal to “space averages”. A draft of the
von Neumann mean ergodic theorem which appeared in 1932 by John von Neumann has
motivated Birkhoff, but the mean ergodic version is weaker. See [439] for history. A special
case is the law of large numbers, in which case the random variables x → X(T k (x)) are
independent with equal distribution (IID). The theorem belongs to ergodic theory [181, 100,
372].
7. Set theory
A bijection is a map from X to Y which is injective: f (x) = f (y) ⇒ x = y and surjective:
for every y ∈ Y , there exists x ∈ X with f (x) = y. Two sets X, Y have the same cardinality,
if there exists a bijection from X to Y . Given a set X, the power set 2X is the set of all
subsets of X, including the empty set and X itself. If X has n elements, the power set has
2n elements. Cantor’s theorem is
Theorem: For any set X, the sets X and 2X have different cardinality.
The result is due to Cantor. Taking for X the natural numbers, then every Y ∈ 2X defines a
real number φ(Y ) = y∈Y 2−y ∈ [0, 1]. As Y and [0, 1] have the same cardinality (as double
P
counting pair cases like 0.39999999 · · · = 0.400000 . . . form a countable set), the set [0, 1]
is uncountable. There are different types of infinities leading to countable infinite sets
and uncountable infinite sets. For comparing sets, the Schröder-Bernstein theorem is
important. If there exist injective functions f : X → Y and g : Y → X, then there exists a
bijection X → Y . This result was used by Cantor already. For literature, see [182].
8. Statistics
A probability space (Ω, A, P) consists of a set Ω, a σ-algebra A and a probability measure
P. A σ-algebra is a collection of subset of Ω which contains the empty set and which is closed
under the operations of taking complements, countable unions and countable S intersections.
P The
function P on A takes values in the interval [0, 1], satisfies P[Ω] = 1 and P[ A∈S A] = A∈S P[A]
3
for any finite or countable set S ⊂ A of pairwise disjoint sets. The elements in A are called
events. Given two events A, B where B satisfies P[B] > 0, one can define the conditional
probability P[A|B] = P[A ∩ B]/P[B]. Bayes theorem states:
Theorem: P[A|B] = P[B|A]P[A]/P[B]
The setup stated the Kolmogorov axioms by Andrey Kolmogorov who wrote in 1933 the
“Grundbegriffe der Wahrscheinlichkeitsrechnung” [259] based on measure theory built by Emile
Borel and Henry Lebesgue. For history, see [362], who report that “Kolmogorov sat down
to write the Grundbegriffe, in a rented cottage on the Klyaz’ma River in November 1932”.
Bayes theorem is more like a fantastically clever definition and not really a theorem. There
is nothing to prove as multiplying with P[B] gives P[A ∩ B] on both sides. It essentially
restates that A ∩ B = B ∩ A, the Abelian property of the product in the ring A. More
general is the statement that
P if A1 , . . . , An is a disjoint set of events whose union is Ω, then
P[Ai |B] = P[B|Ai ]P[Ai ]/( j P[B|Aj ]P[Aj ]. Bayes theorem was first proven in 1763 by Thomas
Bayes. It is by some considered to the theory of probability what the Pythagoras theorem is to
geometry. If one measures the ratio applicability over the difficulty of proof, then this theorem
even beats Pythagoras, as no proof is required. Similarly as “a+(b+c)=(a+b)+c”, also Bayes
theorem is essentially a definition but less intuitive as “Monty Hall” illustrates [347]. See [248].
9. Graph theory
A finite simple graph G = (V, E) is a finite collection V of vertices connected by a finite
collection E of edges, which are un-ordered pairs (a, b) with a, b ∈ V . Simple means that no
self-loops nor multiple connections are present in the graph. The vertex degree d(x) of
x ∈ V is the number of edges containing x.
P
Theorem: x∈V d(x)/2 = |E|.
This formula is also called the Euler handshake formula because every edge in a graph
contributes exactly two handshakes. It can be seen as a Gauss-Bonnet formula for the
valuation G → v1 (G) counting the number of edges in G. A valuation φ is a function
defined on subgraphs with the property that φ(A ∪ B) = φ(A) + φ(B) − φ(A ∩ B). Examples
of valuations are the number vk (G) of complete sub-graphs of dimension k of G. An other
example is the Euler characteristic χ(G) = v0 (G)−v1 (G)+v2 (G)−v3 (G)+· P · ·+(−1)d vd (G).
If we write dk (x) = vk (S(x)), where S(x) is the unit sphere of x, then x∈V dk (x)/(k + 1) =
vk (G) is the generalized handshake formula, the Gauss-Bonnet P result kfor vk . The Euler
characteristic then satisfies x∈V K(x) = χ(G), where K(x) = ∞
P
k=0 (−1) vk (S(x))/(k + 1).
This is the discrete Gauss-Bonnet result. The handshake result was found by Euler. For
more about graph theory, [46, 298, 28, 174] about Euler: [145].
10. Polyhedra
A finite simple graph G = (V, E) is given by a finite vertex set V and edge set E. A subset
W of V generates the sub-graph (W, {{a, b} ∈ E | a, b ∈ W }). The unit sphere of v ∈ V
is the sub graph generated by S(x) = {y ∈ V | {x, v} ∈ E}. The empty graph 0 = (∅, ∅)
is called the (−1)-sphere. The 1-point graph 1 = ({1}, ∅) = K1 is the smallest contractible
graph. Inductively, a graph G is called contractible, if it is either 1 or if there exists x ∈ V
such that both G − x and S(x) are contractible. Inductively, a graph G is called a d-sphere,
4
OLIVER KNILL
if it is either 0 or if every S(x) is a (d − 1)-sphere and if there exists a vertex x such that
G − x is contractible. Let vk denote the number of complete sub-graphs Kk+1 of G. The vector
(v0 , v1 , . . . ) is the f -vector of G and χ(G) = v0 − v1 + v2 − . . . is the Euler characteristic of
G. The generalized Euler gem formula due to Schläfli is:
Theorem: For d = 2, χ(G) = v − e + f = 2. For d-spheres, χ(G) = 1 + (−1)d .
Convex Polytopes were studied already in ancient Greece. The Euler characteristic relations
were discovered in dimension 2 by Descartes [4] and interpreted topologically by Euler who
proved the case d = 2. This is written as v − e + f = 2, where v = v0 , e = v1 , f = v2 . The two
dimensional case can be stated for planar graphs, where one has a clear notion of what the
two dimensional cells are and can use the topology of the ambient sphere in which the graph
is embedded. Historically there had been confusions [89, 342] about the definitions. It was
Ludwig Schläfli [359] who covered the higher dimensional case. The above set-up is a modern
reformulation of his set-up, due essentially to Alexander Evako. Multiple refutations [271] can
be blamed to ambiguous definitions. Polytopes are often defined through convexity [176, 438]
and there is not much consensus on a general definition [175], which was the reason in this entry
to formula Schläfli’s theorem using here a maybe a bit restrictive (as all cells are simplices),
but clear combinatorial definition of what a “sphere” is.
11. Topology
The Zorn lemma assures that that the Cartesian product of a non-empty family of non-empty
sets is non-empty. The Zorn lemma is equivalent to the axiom of choice Q
in the ZFC axiom
system and to the Tychonov theorem in topology as below. Let X = i∈I Xi denote the
product of topological spaces. The product topology is the weakest topology on X which
renders all projection functions πi : X → Xi continuous.
Q
Theorem: If all Xi are compact, then i∈I Xi is compact.
Zorn’s lemma is due to Kazimierz Kuratowski in 1922 and Max August Zorn in 1935. Andrey
Nikolayevich Tykhonov proved his theorem in 1930. One application of the Zorn lemma is the
Hahn-Banach theorem in functional analysis, the existence of spanning trees in infinite
graphs or the fact that commutative rings with units have maximal ideals. For literature, see
[219].
12. Algebraic geometry

The algebraic set V (J) of an ideal J in the commutative ring R = k[x1 , . . . , xn ] over an
algebraically closed field
√ k defines the ideal I(V (J)) containing all polynomials that vanish
on V (J). The radical J of an ideal J is the set of polynomials in R such that rn ∈ J for
some positive n. (An ideal J in a ring R is a subgroup of the additive group of R such that
rx ∈ I for all r ∈ R and all x ∈ I. It defines the quotient ring R/I and is so the kernel of a
ring homomorphism from R to R/I. The algebraic set V (J) = {x ∈ k n | f (x) = 0, ∀f ∈ J}
of an ideal J in the polynomial ring R is the set of common roots of all these functions f .
The algebraic sets are the closed sets in the Zariski topology of R. The ring R/I(V ) is the
coordinate ring of the algebraic set V .) The Hilbert Nullstellensatz is
5
√
Theorem: I(V (J)) = J.
The theorem is due to Hilbert. A simple example is when J = hpi = hx2 − 2xy + y 2 i is the
ideal J generated by p in R[x, y]; then V (J) = {x = y} and I(V (J)) is the ideal generated by
x − y. For literature, see [189].
13. Cryptology
An integer p > 1 is prime if 1 and p are the only factors of p. The number k mod p is the
reminder when dividing k by p. Fermat’s little theorem is
Theorem: ap = a mod p for every prime p and every integer a.
The theorem was found by Pierre de Fermat in 1640. A first proof appeared in 1683 by
Leibniz. Euler in 1736 published the first proof. The result is used in the Diffie-Helleman
key exchange, where a large public prime p and a public base value a are taken. Ana chooses
a number x and publishes X = ax modp and Bob picks y publishing Y = ay modp. Their secret
key is K = X y = Y x . An adversary Eve who only knows a, p, X and Y can from this not get
K due to the difficulty of the discrete log problem. More generally, for possibly composite
numbers n, the theorem extends to the fact that aφ(n) = 1 modulo p, where the Euler’s totient
function φ(n) counts the number of positive integers less than n which are coprime to n. The
generalized Fermat theorem is the key for RSA crypto systems: in order for Ana and Bob
to communicate. Bob publishes the product n = pq of two large primes as well as some base
integer a. Neither Ana nor any third party Eve do know the factorization. Ana communicates a
message x to Bob by sending X = ax modn using modular exponentiation. Bob, who knows
p, q, can find y such that xy = 1 mod φ(n). This is because of Fermat a(p−1)(q−1) = a mod n.
Now, he can compute x = y −1 mod φ(n). Not even Ana herself could recover x from X.
14. Spectral theorem

A bounded linear operator A on a Hilbert space is called normal if AA∗ = A∗ A, where
T
A∗ = A is the adjoint and AT is the transpose and A is the complex conjugate. Examples
of normal operators are self-adjoint operators (meaning A = A∗ ) or unitary operators
(meaning AA∗ = 1).
Theorem: A is normal if and only if A is unitarily diagonalizable.
In finite dimensions, any unitary U diagonalizing A using B = U ∗ AU contains an orthonormal

eigenbasis of A as column vectors. The theorem is due to Hilbert. In the self-adjoint case,
all the eigenvalues are real and in the unitary case, all eigenvalues are on the unit circle. The
result allows a functional calculus for normal operators: for any continuous function f and
any bounded linear operator A, one can define f (A) = U f (B)U ∗ , if B = U ∗ AU . See [94].
15. Number systems

A monoid is a set X equipped with an associative operation ∗ and an identity element
1 satisfying 1 ∗ x = x for all x ∈ X. Associativity means x ∗ (y ∗ z) = (x ∗ y) ∗ z for all
x, y, z ∈ X. The monoid structure belongs to a collection of mathematical structures magmas
⊃ semigroups ⊃ monoids ⊃ groups. A monoid is commutative, if x ∗ y = y ∗ x for
6
OLIVER KNILL
all x, y ∈ X. A group is a monoid in which every element x has an inverse y satisfying

x ∗ y = y ∗ x = 1.
Theorem: Every commutative monoid can be extended to a group.
The general result is due to Alexander Grothendieck from around 1957. The group is called
the Grothendieck group completion of the monoid. For example, the additive monoid of
natural numbers can be extended to the group of integers, the multiplicative monoid of non-zero
integers can be extended to the group of rational numbers. The construction of the group is
used in K-theory [22, 228] For insight about the philosophy of Grothendieck’s mathematics,
see [296].
16. Combinatorics
Let |X| denote the cardinality of a finite set X. This means that |X| is the number of elements
in X. A function f from a set X to a set Y is called injective if f (x) = f (y) implies x = y.
The pigeonhole principle tells:
Theorem: If |X| > |Y | then no function X → Y can be injective.
This implies that if we place n items into m boxes and n > m, then one box must contain
more than one item. The principle is believed to be formalized first by Peter Dirichlet. Despite
its simplicity, the principle has many applications, like proving that something exists. An
example is the statement that there are two trees in New York City streets which have the
same number of leaves. The reason is that the U.S. Forest services states 592’130 trees in
the year 2006 and that a mature, healthy tree has about 200’000 leaves. One can also use
it for less trivial statements like that in a cocktail party there are at least two with the same
number of friends present at the party. A mathematical application is the Chinese remainder
Theorem stating that that there exists a solution to ai x = bi mod mi all disjoint pairs mi , mj
and all pairs ai , mi are relatively prime [116, 289]. The principle generalizes to infinite set if
|X| is the cardinality. It implies then for example that there is no injective function from the
real numbers to the integers. For literature, see for example [64], which states also a stronger
version which for example allows to show that any sequence of real n2 + 1 real numbers contains
either an increasing subsequence of length n + 1 or a decreasing subsequence of length n + 1.
17. Complex analysis
Assume f is an analytic function in an open domain G of the complex plane C. Such
a function is also called holomorphic in G. Holomorphic means that if f (x + iy) = u(x +
iy) + iv(x + iy), then the Cauchy-Riemann differential equations ux = vy , uy = −vx hold in
G. Assume z is in G and assume C ⊂ G is a circle z + reiθ centered at z which is bounding a
disc D = {w ∈ C | |w − z| < r} ⊂ G.
Theorem: For analytic f in G and a circle C ⊂ G, one has f (w) = C f(z−w)

R (z)dz
.
This Cauchy integral formula of Cauchy is used for other R results and estimates. It implies
for example the Cauchy integral theorem assuring that C f (z)dz = 0 for any simple closed
curve C in G boundingR a simply connected region D ⊂ G. Morera’s theorem assures that
for any domain G, if C f (z) dz = 0 for all simple closed smooth curves C in G, then f is
holomorphic in G. An other generalization is residue calculus: For a simply connected region
7
G and a function f which is analytic except in a Rfinite set A of points.

P If C is piecewise smooth
continuous closed curve not intersecting A, then C f (z) dz = 2πi a∈A I(C, a)Res(f, a), where
I(C, a) is the winding number of C with respect to a and Res(f, a) is the residue of f at a
which is in the case of poles given by limz→a (z − a)f (z). See [75, 8, 93].
18. Linear algebra

If A is a m × n matrix with image ran(A) and kernel ker(A). If V is a linear subspace of
Rm , then V ⊥ denotes the orthogonal complement of V in Rm , the linear space of vectors
perpendicular to all x ∈ V .
Theorem: dim(kerA) + dim(ranA) = n, dim((ranA)⊥ ) = dim(kerAT ).
The result is used in data fitting for example when understanding the least square solu-
tion x = (AT A)−1 AT b of a system of linear equations Ax = b. It assures that AT A is
invertible if A has a trivial kernel. The result is a bit stronger than the rank-nullity theorem
dim(ran(A)) + dim(ker(A)) = n alone and implies that for finite m × n matrices the index
dim(kerA) − dim(kerA∗ ) is always n − m, which is the value for the 0 matrix. For literature, see
[393]. The result has an abstract generalization in the form of the group isomorphism theorem
for a group homomorphism f stating that G/ker(f ) is isomorphic to f (G). It can also be
described using the singular value decomposition A = U DV T . The number r = ranA has
as a basis the first r columns of U . The number n − r = kerA has as a basis the last n − r
columns of V . The number ranAT has as a basis the first r columns of V . The number kerAT
has as a basis the last m − r columns of U .
19. Differential equations

A differential equation dtd x = f (x) and x(0) = x0 in a Banach space (X, || · ||) (a normed,
complete vector space) defines an initial value problem: we look for a solution x(t) satisfying
the equation and given initial condition x(0) = x0 and t ∈ (−a, a) for some a > 0. A function
f from R to X is called Lipschitz, if there exists a constant C such that for all x, y ∈ X the
inequality ||f (x) − f (y)|| ≤ C|x − y| holds.
Theorem: If f is Lipschitz, a unique solution of x0 = f (x), x(0) = x0 exists.
This result is due to Picard and Lindelöf from 1894. Replacing the Lipschitz condition with
continuity still gives an existence√ theorem which is due to Giuseppe Peano in 1886, but
uniqueness can fail like for x0 = x, x(0) = 0 with solutions x = 0 and x(t) = t2 /4. The
example x0 (t) = x2 (t), x(0) = 1 with solution 1/(1 − t) shows that we can not have solutions
for all t. The proof is a simple application of the Banach fixed point theorem. For literature,
see [88].
20. Logic
An axiom system A is a collection of formal statements assumed to be true. We assume it to
contain the basic Peano axioms of arithmetic. An axiom system is complete, if every true
statement can be proven within the system. The system is consistent if one can not prove
1 = 0 within the system. It is provably consistent if one can prove a theorem ”The axiom
system A is consistent.” within the system.
8
OLIVER KNILL
Theorem: An axiom system is neither complete nor provably consistent.

The result is due to Kurt Goedel who proved it in 1931. In this thesis, Goedel had proven
a completeness theorem of first order predicate logic. The incompleteness theorems of 1931
destroyed the dream of Hilbert’s program which aimed for a complete and consistent axiom
system for mathematics. A commonly assumed axiom system is the Zermelo-Frenkel axiom
system together with the axiom of choice ZFC. Other examples are Quine’s new foundations
NF or Lawvere’s elementary theory of the category of sets ETCS. For a modern view on
Hilbert’s program, see [402]. For Goedel’s theorem [151, 312]. Hardly any other theorem had
so much impact outside of mathematics.
21. Representation theory

For a finite group or compact topological group G, one can look at representations,
group homomorphisms from G to the automorphisms of a vector space V . A representation
of G is irreducible if the only G-invariant subspaces of V are 0 or V . The direct sum of of
two representations φ, ψ is defined as φ ⊕ ψ(g)(v ⊕ w) = φ(g)(v) ⊕ φ(g)(w). A representation is
semi simple if it is a unique direct sum of irreducible finite-dimensional representations:
Theorem: Representations of compact topological groups are semi simple.
For representation theory, see [425]. Pioneers in representation theory were Ferdinand Georg
Frobenius, Herman Weyl, and Élie Cartan. Examples of compact groups are finite group, or
compact Lie groups (a smooth manifold which is also a group for which the multiplications
and inverse operations are smooth) like the torus group T n , the orthogonal groups O(n) of
all orthogonal n × n matrices or the unitary groups U (n) of all unitary n × n matrices or the
group Sp(n) of all symplectic n × n matrices. Examples of groups that are not Lie groups are
the groups Zp of p-adic integers, which are examples of pro-finite groups.
22. Lie theory

Given a topological group G, a Borel measure µ on G is called left invariant if µ(gA) =
µ(A) for every g ∈ G and every measurable set A ⊂ G. A left-invariant measure on G is also
called a Haar measure. A topological space is called locally compact, if every point has a
compact neighborhood.
Theorem: A locally compact group has a unique Haar measure.
Alfréd Haar showed the existence in 1933 and John von Neumann proved that it is unique.
In the compact case, the measure is finite, leading to an inner product and so to unitary
representations. Locally compact Abelian groups G can be understood by their characters,
continuous group homomorphisms from G to the circle group T = R/Z. The set of characters
defines a new locally compact group Ĝ, the dual of G. The multiplication is the pointwise
multiplication, the inverse is the complex conjugate and the topology is the one of uniform
convergence on compact sets. If G is compact, then Ĝ is discrete, and if G is discrete, then Ĝ
is compact. In orderRto prove Pontryagin duality ˆˆG = G, one needs a generalized Fourier
transform fˆ(χ) = G f (x)χ(x)dµ(x) which uses the Haar measure. The inverse Fourier
transform gives back f using the Rdual Haar measure. The Haar measure is also used to
define the convolution f ? g(x) = G f (x − y)g(y)dµ(y) rendering L1 (G) a Banach algebra.
9
The Fourier transform then produces a homomorphism from L1 (G) to C0 (Ĝ) or a unitary
transformation from L2 (G) to L2 (Ĝ). For literature, see [83, 416].
23. Computability
The class of general recursive functions is the smallest class of functions which allows
projection, iteration, composition and minimization. The class of Turing computable
functions are the functions which can be implemented by a Turing machine possessing
finitely many states. Turing introduced this in 1936 [327].
Theorem: The generally recursive class is the Turing computable class.

Kurt Goedel and Jacques Herbrand defined the class of general recursive functions around 1933.
They were motivated by work of Alonzo Church who then created λ calculus later in 1936.
Alan Turing developed the idea of a Turing machine which allows to replace Herbrand-Goedel
recursion and λ calculus. The Church thesis or Church-Turing thesis states that everything
we can compute is generally recursive. As “whatever we can compute” is not formally defined,
this always will remain a thesis unless some more effective computation concept would emerge.
24. Category theory

Given an element A in a category C, let hA denote the functor which assigns to a set X the
set Hom(A, X) of all morphisms from A to X. Given a functor F from C to the category
S = Set, let N (G, F ) be the set of natural transformations from G = hA to F . (A natural
transformation between two functors G and F from C to S assigns to every object x in
C a morphism ηx : G(x) → F (x) such that for every morphism f : x → y in C we have
ηy ◦ G(f ) = F (f ) ◦ ηx .) The functor category defined by C and S has as objects the functors
F and as morphisms the natural transformations. The Yoneda lemma is
Theorem: N (hA , F ) can be identified with F (A).
Category theory was introduced in 1945 by Samuel Eilenberg and Sounders Mac Lane. The
lemma above is due to Nobuo Yoneda from 1954. It allows to see a category embedded in a
functor category which is a topos and serves as a sort of completion. One can identify a
set S for example with Hom(1, S). An other example is Cayley’s theorem stating that the
category of groups can be completely understood by looking at the group of permutations of
G. For category theory, see [295, 272]. For history, [267].
25. Perturbation theory

A function f of several variables is called smooth if one can take first partial derivatives
like ∂x , ∂y and second partial derivatives like ∂x ∂y f (x, y) = fxy (x, y) and still get continuous
function. Assume f (x, y) is a smooth function of two Euclidean variables x, y ∈ Rn . If
f (a, 0) = 0, we say a is a root of x → f (x, y). If fy (x0 , y) is invertible, the root is called
non-degenerate. If there is a solution f (g(y), y) = 0 such that g(0) = a and g is continuous,
the root a has a local continuation and say that it persists under perturbation.
Theorem: A non-degenerate root persists under perturbation.
10
OLIVER KNILL
This is the implicit function theorem. There are concrete and fast algorithms to compute the
continuation. An example is the Newton method which iterates T (x) = x − f (x, y)/fx (x, y)
to find the roots of x → f (x, y) for fixed y. The importance of the implicit function theorem
is both theoretical as well as applied. The result assures that one can makes statements about
a complicated theory near some model, which is understood. There are related situations, like
if we want to continue a solution of F (x, y) = (f (x, y), g(x, y)) = (0, 0) giving equilibrium
points of the vector field F . Then the Newton step T (x, y) = (x, y) − dF −1 (x, y) · F (x, y)
method allows a continuation if dF (x, y) is invertible. This means that small deformations of
F do not lead to changes of the nature of the equilibrium points. When equilibrium points
change, the system exhibits bifurcations. This in particular applies to F (x, y) = ∇f (x, y),
where equilibrium points are critical points. The derivative dF of F is then the Hessian.
26. Counting
A simplicial complex X is a finite set of non-empty sets that is closed under the operation
of taking finite non-empty subsets. The Euler characteristic χ of a simplicial complex G is
defined as χ(X) = x∈X (−1)dim(x) , where the dimension dim(x) of a set x is its cardinality
P
|x| minus 1.
Theorem: χ(X × Y ) = χ(X)χ(Y ).
For zero-dimensional simplicial complexes G, (meaning that all sets in G have cardinality
1), we get the rule of product: if you have m ways to do one thing and n ways to do
an other, then there are mn ways to do both. This fundamental counting principle is
used in probability theory for example. The Cartesian product X × Y of two complexes
is defined as the set-theoretical product of the two finite sets. It is not a simplicial complex
any more in general but has the same Euler characteristic than its Barycentric refinement
(X × Y )1 , which
Pnis a simplicial complex. The maximal dimension of A × B is dim(A) + dim(B)
k
and pX (t) = v
k=0 k (X)t is the generating function of vk (X), then pX×Y (t) = pX (t)pY (t)
implying the counting principle as pX (−1) = χ(X). The function pX (t) is called the Euler
polynomial of X. The importance of Euler characteristic as a counting tool lies in the fact
that only χ(X) = pX (−1) is invariant under Barycentric subdivision χ(X) = X1 , where X1
is the complex which consists of the vertices of all complete subgraphs of the graph in which
the sets of X are the vertices and where two are connected if one is contained in the other.
The concept of Euler characteristic goes so over to continuum spaces like manifolds where the
product property holds too. See for example [11].
27. Metric spaces

A continuous map T : X → X, where (X, d) is a complete non-empty metric space is called
a contraction if there exists a real number 0 < λ < 1 such that d(T (x), T (y)) ≤ λd(x, y) for
all x, y ∈ X. The space is called complete if every Cauchy sequence in X has a limit. (A
sequence xn in X is called Cauchy if for all > 0, there exists n > 0 such that for all i, j > n,
one has d(xi , xj ) < .)
Theorem: A contraction has a unique fixed point in X.

This result is the Banach fixed point theorem proven by Stefan Banach from 1922. The
example case T (x) = (1 − x2 )/2 on X = Q ∩ [0.3, 0.6] having contraction rate λ = 0.6 and
11
T (X) = Q√ ∩ [0.32, 0.455] ⊂ X shows that completeness √ is necessary. The unique fixed point of
T in X is 2 − 1 = 0.414... which is not in Q because 2 = p/q would imply 2q 2 = p2 , which
is not possible for integers as the left hand side has an odd number of prime factors 2 while the
right hand side has an even number of prime factors. See [323]
28. Dirichlet series

The abscissa of simple convergence of a Dirichlet series ζ(s) = ∞ −λn s
P
n=1 an e is σ0 =
inf{a ∈ P R | ζ(z) converges for all Re(z) > a }. For λn = n we have the Taylor series
f (z) = ∞ n −s
n=1 an z with z = e . For λn = log(n) we have the standard Dirichlet P∞ n series
P ∞ s n s
a
n=1 n /n . For example, for a n = z , one gets the poly-logarithm Lis (z) = n=1 z /n and
especially
P∞ n Lis (1) = ζ(s), the RiemannPn zeta function or the Lerch transcendent Φ(z, s, a) =
s
n=1 z /(n + a) . Define S(n) = k=1 ak . The Cahen’s formula applies if the series S(n)
does not converge.
log |S(n)|
Theorem: σ0 = lim supn→∞ λn
.
There is a similar formula for the abscissa of absolute convergence of ζ which is defined
as σa = inf{a ∈ R | ζ(z) converges absolutely for all Re(z) > a }. The result is σa =
P∞
lim supn→∞ log(S(n))
λn
, For example, for the Dirichlet eta function ζ(s) = n=1 (−1)
n−1
/ns
has the abscissa
P∞of convergence σ0 = 0 and the absolute abscissa of convergence σa = 1. The
inα s
series ζ(s) = n=1 e /n has P σa = 1 sandQ σ0 = 1 − α. If an is multiplicative an+m = an am for
relatively prime n, m, then ∞ a
n=1 Qn /n = s 2s
p (1 + ap /p + ap2 /p + · · · ) generalizes the Euler
golden key formula n 1/ns = p (1 − 1/ps )−1 . See [184, 186].
P
29. Trigonometry
Mathematicians had a long and painful struggle with the concept of limit. One of the first
to ponder the question was Zeno of Elea around 450 BC. Archimedes of Syracuse made some
progress around 250 BC. Since Augustin-Louis Cauchy, one defines the limit limx→a f (x) = b to
exist if and only if for all > 0, there exists a δ > 0 such that if |x − a| < δ, then |f (x) − b| < .
A place where limits appear are when computing derivatives g 0 (0) = limx→0 [g(x)−g(0)]/x. In
the case g(x) = sin(x), one has to understand the limit of the function f (x) = sin(x)/x which is
the sinc function. A prototype result is the fundamental theorem of trigonometry (called
as such in some calculus texts like [61]).
Theorem: limx→0 sin(x)/x = 1.
It appears strange to give weight to such a special result but it explains the difficulty of limit
and the l’Hôpital rule of 1694, which was formulated in a book of Bernoulli commissioned to
Hôpital: the limit can be obtained by differentiating both the denominator and nominator and
taking the limit of the quotients. The result allows to derive (using trigonometric identities)
that in general sin0 (x) = cos(x) and cos0 (x) = − sin(x). One single limit is the gateway. It is im-
portant also culturally because it embraces thousands of years of struggle. It was Archimedes,
who used the theorem when computing the circumference of the circle formula 2πr using
exhaustion using regular polygons from the inside and outside. Comparing the lengths of
12
OLIVER KNILL
the approximations essentially battled that fundamental theorem of trigonometry. The iden-
tity is therefore the epicenter around the development of trigonometry, differentiation and
integration.
30. Logarithms
The natural logarithm is the inverse of the exponential function exp(x) establishing so a
group homomorphism from the additive group (R, +) to the multiplicative group (R+ , ∗).
We have:
Theorem: log(uv) = log(u) + log(v).
This follows from exp(x + y) = exp(x) exp(y) and log(exp(x)) = exp(log(x)) = x by plugging
in x = log(u), y = log(v). The logarithms were independently discovered by Jost Bürgi around
1600 and John Napier in 1614 [382]. The logarithm with base b > 0 is denoted by logb . It is the
inverse of x → bx = ex log(b) . The concept of logarithm has been extended in various ways: in any
group G, one can define the discrete logarithm logb (a) to base b as an integer k such that
bk = a (if it exists). For complex numbers the complex logarithm log(z) as any solution w of
ew = z. It is multi-valued as log(|z|) + iarg(z) + 2πik all solve this with some integer k, where
arg(z) ∈ (−π, π). The identity log(uv) = log(u)+log(v) is now only true up to 2πki. Logarithms
can also be defined for matrices. Any matrix B solving exp(B) = A is called a logarithm of
A. For A close to the identity I, can define log(A) = (A − I) − (A − I)2 /2 + (A − I)3 /3 − ...,
which is a Mercator series. For normal invertible matrices, one can define logarithms
using the functional calculus by diagonalization. On a Riemannian manifold M , one also
has an exponential map: it is a diffeomorphim from a small ball Br (0) in the tangent space
x ∈ M to M . The map v → expx (v) is obtained by defining expx (0) = x and by taking for
v 6= 0 a geodesic with initial direction v/|v| and running it for time |v|. The logarithm logx
is now defined on a geodesic ball of radius r and defines an element in the tangent space. In
the case of a Lie group M = G, where the points are matrices, each tangent space is its Lie
algebra.
31. Geometric probability

A subset K of Rn is called compact if it is closed and bounded. By Bolzano-Weierstrass
this is equivalent to the fact that every infinite sequence xn in K has a subsequence which
converges. A subset K of Rn is called convex, if for any two given points x, y ∈ K, the
interval {x + t(y − x), t ∈ [0, 1]} is a subset of K. Let G be the set of all compact convex
subsets of Rn . An invariant valuation X is a function X : G → R satisfying X(A ∪
B) + X(A ∩ B) = X(A) + X(B), which is continuous in the Hausdorff metric d(K, L) =
max(supx∈K inf y∈L d(x, y) + supy∈K inf x∈L d(x, y)) and invariant under rigid motion generated
by rotations, reflections and translations in the linear space Rn .
Theorem: The space of valuations is (n + 1)-dimensional.
The theorem is P
due to Hugo Hadwiger from 1937. The coefficients aj (G) of the polynomial
Vol(G + tB) = nj=0 aj tj are a basis, where B is the unit ball B = {|x| ≤ 1}. See [239].
13
32. Partial differential equations

A quasilinear partial differential equation is a differential equation of the form ut (x, t) =
F (x, t, u) · ∇x u(x, t) + f (x, t, u) with analytic initial condition u(x, 0) = u0 (x) and an analytic
vector field F . It defines a quasi-linear Cauchy problem.
Theorem: A quasi-linear Cauchy problem has a unique analytic solution.
This is the Cauchy-Kovalevskaya theorem. It was initiated by Augustin-Louis Cauchy
in 1842 and proven in 1875 by Sophie Kowalevskaya. Analyticity is important, smoothness
alone is not enough. If F is analytic in each variable, one can look at equations like the
Cauchy problem ut = F (t, x, u, ux , uxx ). Examples are partial differential equations like the heat
equation ut = uxx or the wave equation utt = uxx . Given an initial condition u(0, x) = u0 (x)
one then deals with an ordinary differential equation in a function space. One can then try
to approach the Cauchy-Kovalevskaya problem by Picard-Lindelöf. The problem is that the
Lipschitz condition fails because the corresponding operators are unbounded. Even Cauchy-
Peano (which does not ask for uniqueness) fails. And this even in an analytic setting. [317]
gives the example ut = uxx with initial condition u(0, x) = 1/(1 + x2 ) for which the entire series
solving the problem has a zero radius of convergence in x for any t > 0. Texts like [403, 317]
give full versions of the Cauchy-Kovalevskaya theorem for real-analytic Cauchy initial data on
a real analytic hypersurface satisfying a non-characteristic condition for the partial differential
equation. For a shorter introduction to partial differential equations, see [19].
33. Game theory

If S = (S1 , . . . , Sn ) are n players and f = (f1 , . . . , fn ) is a payoff function defined on a
strategy profile x = (x1 , . . . , xn ). A point x∗ is called an equilibrium if fi (x∗ ) is maximal
with respect to changes of xi alone in the profile x for every player i.
Theorem: There is an equilibrium for any game with mixed strategy
The equilibrium is called a Nash equilibrium. It tells us what we would see in a world if
everybody is doing their best, given what everybody else is doing. John Forbes Nash used
in 1950 the Brouwer fixed point theorem and later in 1951 the Kakutani fixed point
theorem to prove it. The Brouwer fixed point theorem itself is generalized by the Lefschetz
fixed point theorem which equates the super trace of the induced map on cohomology with
the sum of the indices of the fixed points. About John Nash and some history of game theory,
see [365]: game theory started maybe with Adam Smith’s the Wealth of Nations published
in 1776, Ernst Zermelo in 1913 (Zermelo’s theorem), Émile Borel in the 1920s and John von
Neumann in 1928 pioneered mathematical game theory. Together with Oskar Morgenstern,
John von Neumann merged game theory with economics in 1944. Nash published his thesis in
a paper of 1951. For the mathematics of games, see [419].
34. Measure theory

A topological space with open sets O defines the Borel σ-algebra, the smallest σ algebra
which contains O. For the metric space (R, d) with d(x, y) = |x − y|, already the intervals
generate the Borel σ algebra A. A Borel measure is a measure defined on a Borel σ-algebra.
Every Borel measure µ on the real line R can be decomposed uniquely into an absolutely
continuous part µac , a singular continuous part µsc and a pure point part µpp :
14
OLIVER KNILL
Theorem: µ = µac + µsc + µpp .
This is called the Lebesgue decomposition theorem. It uses the Radon-Nikodym the-
orem. The decomposition theorem implies the decomposition theorem of the spectrum of
a linear operator. See [371] (like page 259). Lebesgue’s theorem was published in 1904. A
generalization due to Johann Radon and Otto Nikodym was done in 1913.
35. Geometric number theory

If Γ is a lattice in Rn , denote with Rn /Γ the fundamental region and by |Γ| its volume. A
set K is convex if x, y ∈ K implies x + t(x − y) ∈ K for all 0 ≤ t ≤ 1. A set K is centrally
symmetric if x ∈ K implies −x ∈ K. A region is Minkowski if it is convex and centrally
symmetric. Let |K| denote the volume of K.
Theorem: If K is Minkowski and |K| > 2n |Γ| then K ∩ Γ 6= ∅.
The theorem is due to Hermann Minkowski in 1896. It lead to a field called geometry of
numbers. [77]. It has many applications in number theory and Diophantine analysis
[68, 204]
36. Fredholm
An integral kernel K(x, y) ∈ L2 ([a, b]2 ) defines an integral operator A defined by Af (x) =
Rb Rb
a
K(x, y)f (y) dy with adjoint T ∗ f (x) = a K(y, x)f (y) dy. The L2 assumption makes the
function K(x, y) what one calls a Hilbert-Schmidt kernel. Fredholm showed that the Fred-
holm equation A∗ f = (T ∗ − λ)f = g has a solution f if and only if f is perpendicular to
the kernel of A = T − λ. This identity ker(A)⊥ = im(A∗ ) is in finite dimensions part of the
fundamental theorem of linear algebra. The Fredholm alternative reformulates this in
a more catchy way as an alternative:
Theorem: Either ∃f 6= 0 with Af = 0 or for all g, ∃f with Af = g.
In the second case, the solution depends continuously on g. The alternative can be put more
generally by stating that if A is a compact operator on a Hilbert space and λ is not an
eigenvalue of A, then the resolvent (A − λ)−1 is bounded. A bounded operator A on a Hilbert
space H is called compact if the image of the unit ball is relatively compact (has a compact
closure). The Fredholm alternative is part of Fredholm theory. It was developed by Ivar
Fredholm in 1903.
37. Prime distribution

The Dirichlet theorem about the primes along an arithmetic progression tells that if a and b
are relatively prime meaning that there largest common divisor is 1, then there are infinitely
many primes of the form p = a mod b. The Green-Tao theorem strengthens this. We say
that a set A contains arbitrary long arithmetic progressions if for every k there exists an
arithmetic progression {a + bj, j = 1, · · · , k} within A.
Theorem: The set of primes contains arbitrary long arithmetic progressions.
15
The Dirichlet prime number theorem was found in 1837. The Green-Tao theorem
was done in 2004 and appeared in 2008 [171]. It uses Szemerédi’s theorem [154] which
shows that any set A of positive upper density lim supn→∞ |A ∩ {1 · · · n}|/n has arbitrary long
arithmetic progressions. So, any subset A of the primes P for which the relative density
lim supn→∞ |A ∩ {1 · · · n}|/|P ∩ {1 · · · n}| is positive has arbitrary long arithmetic progressions.
For non-linear sequences of numbers the problems are wide open. The Landau problem of
the infinitude of primes of the form x2 + 1 illustrates this. The Green-Tao theorem gives hope
to tackle the Erdös P conjecture on arithmetic progressions telling that a sequence {xn }
of integers satisfying n xn = ∞ contains arbitrary long arithmetic progressions.
38. Riemannian geometry

A Riemannian manifold is a smooth finite dimensional manifold M equipped with a sym-
metric, positive definite tensor (u, v) → gx (u, v) defining on each tangent space Tx M an
inner product (u, v)x = (gx (u, v)u, v), where (u, v) is the standard inner product. Let Ω
be the space of smooth vector fields. A connection is a bilinear map (X, Y ) → ∇X Y
from Ω × Ω to Ω satisfying the differentiation rules ∇f X Y = f ∇X Y and Leibniz rule
∇X (f Y ) = df (X)Y + f ∇X Y . It is compatible with the metric if the Lie derivative
satisfies δX (Y, Z) = (ΓX Y, Z) + (Y, ΓX Z). It is torsion-free if ∇X Y − ∇Y X = [X, Y ] is the
Lie bracket on Ω.
Theorem: There is exactly one torsion-free connection compatible with g.
This is the fundamental theorem of Riemannian geometry. The connection is called the
Levi-Civita connection, named after Tullio Levi-Civita. See for example [122, 3, 379, 105].
39. Symplectic geometry

A symplectic manifold (M, ω) is a smooth 2n-manifold M equipped with a non-degenerate
closed 2-form ω. The later is called a symplectic form. As a 2-form, it satisfies ω(x, y) =
−ω(y, x). Non-degenerate P means ω(u, v) = 0 for all v implies u = 0. The standard
symplectic form is ω0 = i<j dxi ∧ dxj .
Theorem: Every symplectic form is locally diffeomorphic to ω0 .

This theorem is due to Jean Gaston Darboux from 1882. Modern proofs use Moser’s trick
from 1965. The Darboux theorem assures that locally, two symplectic manifolds of the same
dimension are symplectic equivalent. It also implies that symplectic matrices have deter-
minant 1. In contrast, for Riemannian manifolds, one can not trivialize the Riemannian
metric in a neighborhood one can only render it the standard metric at the point itself. See
[199].
40. Differential topology

Given a smooth function f on a differentiable manifold M . Let df denote the gradient
of f . A point x is called a critical point, if df (x) = 0. We assume f has only finitely many
critical points and that all of them are non-degenerate. The later means that the Hessian
d2 f (x) is invertible at x. One calls such functions Morse functions. The Morse index
of a critical point x is the number of negative eigenvalues of d2 f . The Morse inequalities
relate the number ck (f, K) of critical points of index k of f with the Betti numbers bk (M ),
16
OLIVER KNILL
defined as the nullity of the Hodge star operator dd∗ + d∗ d restricted to k-forms Ωk , where
dk : Ωk → Ωk+1 is the exterior derivative.
Theorem: ck − ck−1 + · · · + (−1)k c0 ≥ bk − bk−1 + · · · + (−1)k b0 .
These are the Morse inequalities due to Marston Morse from 1934. It implies in particular
the weak Morse inequalities bk ≤ ck . Modern proofs use Witten deformation [105] of the
exterior derivative d.
41. Non-commutative geometry

A spectral triple (A, H, D) is given by a Hilbert space H, a C ∗ -algebra A of operators on
H and a densely defined self-adjoint operator D satisfying ||[D, a]|| < ∞ for all a ∈ A such
2
that e−tD is trace class. The operator D is called a Dirac operator. The set-up generalizes
Riemannian geometry because of the following result dealing with the exterior derivative d
on a Riemannian manifold (M, g), where A = C(M ) is the C ∗ -algebra of continuous functions
and D = d + d∗ is the Dirac operator, defining the spectral triple of (M, g). Let δ denote the
geodesic distance in (M, g):
Theorem: δ(x, y) = supf ∈A,||[D,f ]||≤1 |f (x) − f (y)|.
This formula of Alain Connes tells that the spectral triple determines the geodesic distance
in (M, g) and so the metric g. It justifies to look at spectral triples as non-commutative
generalizations of Riemannian geometry. See [90].
42. Polytopes
A convex polytop P in dimension n is the convex hull of finitely many points in Rn . One
assumes all vertices to be extreme points, points which do not lie in an open line segment
of P . The boundary of P is formed by (n − 1) dimensional boundary facets. The notion
of Platonic solid is recursive. A convex polytop is Platonic, if all its facets are Platonic
(n − 1)-dimensional polytopes and vertex figures. Let p = (p2 , p3 , p4 , . . . ) encode the number
of Platonic solids meaning that pd is the number of Platonic polytops in dimension d.
Theorem: There are 5 platonic solids and p = (∞, 5, 6, 3, 3, 3, . . . )
In dimension 2, there are infinitely many. They are the regular polygons. The list of
Platonic solids is “octahedron”, “dodecahedron”, “icosahedron”, “tetrahedron” and “cube”
has been known by the Greeks already. Ludwig Schläfli first classified the higher dimensional
case. There are six in dimension 4: they are the “5 cell”, the “8 cell” (tesseract), the “16
cell”, the “24 cell”, the “120 cell” and the “600 cell”. There are only three regular polytopes in
dimension 5 and higher, where only the analog of the tetrahedron, cube and octahedron exist.
For literature, see [176, 438, 342].
43. Descriptive set theory

A metric space (X, d) is a set with a metric d (a function X × X → [0, ∞) satisfying
symmetry d(x, y) = d(y, x), the triangle inequality d(x, y) + d(y, z) ≥ d(x, z), and d(x, y) =
0 ↔ x = y.) A metric space (X, d) is complete if every Cauchy sequence converges in X. A
17
metric space is of second Baire category if the intersection of a countable set of open dense
sets is dense. The Baire Category theorem tells
Theorem: Complete metric spaces are of second Baire category.
One calls the intersection A of a countable set of open dense sets A in X also a generic set or
residual set. The complement of a generic set is also called a meager set or negligible or
a set of first category. It is the union of countably many nowhere dense sets. Like measure
theory, Baire category theory allows for existence results. There can be surprises: a generic
continuous function is not differentiable for example. For descriptive set theory, see [237]. The
frame work for classical descriptive set theory often are Polish spaces, which are separable
complete metric spaces. See [56].
44. Calculus of variations

Let X be the vector space of smooth, compactly supported functions h on an interval (a, b).
The fundamental lemma of calculus of variations tells
Rb
Theorem: a
f (x)g(x)dx = 0 for all g ∈ X, then f = 0.
The result is due to Joseph-Louis Lagrange. OneR can restate this as the fact that if f = 0
b
weakly then f is actually zero. It implies that if a f (x)g 0 (x) dx = 0 for all g ∈ X, then f is
constant. This is nice as f is not assumed to be
R b differentiable. The result is used to prove that
0
extrema to a variational problem I(x) = a L(t, x, x ) dt are weak solutions of the Euler
Lagrange equations Lx = d/dtLx0 . See [160, 308].
45. Integrable systems

Given a Hamilton differential equation x0 = J∇H(x) on a compact symplectic 2n-
manifold (M, ω). The almost complex structure J : T ∗ M → T M is tied to ω using a
Riemannian metric g by ω(v, w) = hv, Jgi. A function F : M → R is called an first integral
if d/dtF (x(t)) = 0 for all t. An example is the Hamiltonian function H itself. A set of
integrals F1 , . . . , Fk Poisson commutes if {Fj , Fk } = J∇Fj · ∇Fk = 0 for all k, j. They
are linearly independent, if at every point the vectors ∇Fj are linearly independent in the
sense of linear algebra. A system is Liouville integrable if there are d linearly independent,
Poisson commuting integrals. The following theorem due to Liouville and Arnold characterizes
the level surfaces {F = c} = {F1 = c1 , . . . Fd = cd }:
Theorem: For a Liouville integrable system, level surfaces F = c are tori.
An example how to get integrals is to write the system as an isospectral deformation of

an operator L. This is called a Lax system. Such a differential equation has the form
L0 = [B, L], where B = B(L) is skew symmetric. An example is the periodic Toda system
ȧn = an (bn+1 − bn ), ḃn = 2(a2n − a2n−1 ), where (Lu)n = an un+1 + an−1 un−1 + bn un and (Bu)n =
an un+1 − an−1 un−1 . An other example is the motion of a rigid body in n dimensions if the
center of mass is fixed. See [18].
18
OLIVER KNILL
46. Harmonic analysis

On the vector space X of continuously differentiable 2π periodic, complex- valued functions,
−1
R
define the inner product (f, g) = (2π) f (x)g(x) dx. The Fourier coefficients of f are
ˆ inx
fn = (f, en ), where {en (x) = e }n∈Z is the Fourier basis. The Fourier series of f is the
sum n∈Z fˆn einx .
P
Theorem: The Fourier series of f ∈ X converges point-wise to f .
Already Fourier claimed this always to be true in his “Théorie Analytique de la Chaleur”. After
many fallacious proofs, Dirichlet gave the first proof of convergence [258]. The case is subtle
as there are continuous functions for which the convergence fails at some points. Lipót Féjer
was able to show that for a continuous function f , the coefficients fˆn nevertheless determine
the function using Césaro convergence. See [236].
47. Jensen inequality

If V is a vector space, a set X is called convex if for all points a, b ∈ X, the line segment
{tb + (1 − t)a | t ∈ [0, 1]} is contained in X. A real-valued function φ : X → R is called convex
if φ(tb + (1 − t)a) ≤ tφ(b) + (1 − t)φ(a) for all a, b ∈ X and all t ∈ [0, 1]. Let now R(Ω, A, P) be a
probability space, and f ∈ L1 (Ω, P) an integrable function. We write E[f ] = ω f (x) dP (x)
for the expectation of f . For any convex φ : R → R and f ∈ L1 (Ω, P ), we have the Jensen
inequality
Theorem: φ(E[f ]) ≤ E[φ(f )].
For φ(x) = exp(x) and a finite probability space Ω = {1, 2, . . . , n} with f (k) = xk = exp(yk )
and P[{x}] = 1/n, this gives the arithmetic mean- geometric mean inequality (x1 ·
x2 · · · xn )1/n ≤ (x1 + x2 + · · · + xn )/n. The case φ(x) = ex is useful in general as it leads to the
inequality eE[f ] ≤ E[ef ] if ef ∈ L1 . For f ∈ L2 (ω, P ) one gets (E[f ])2 ≤ E[f 2 ] which reflects the
fact that E[f 2 ] − (E[f ])2 = E[(f − E[f ])2 ] = Var[f ] ≥ 0 where Var[f ] is the variance of f .
48. Jordan curve theorem

A closed curve in the image of a continuous map T → R2 . It is called simple, if this map
is injective. One then calls the map an embedding and the image a topological 1-sphere
or a Jordan curve. The Jordan curve theorem deals with simple closed curves S in the
two-dimensional plane.
Theorem: A simple closed curve divides the plane into two regions.
The Jordan curve theorem is due to Camille Jordan. His proof [222] was objected at first [241]
but rehabilitated in [179]. The theorem can be strengthened, a theorem of Schoenflies tells
that each of the two regions is homeomorphic to the disk {(x, y) ∈ R2 | x2 + y 2 < 1}. In the
smooth case, it is even possible to extend the map to a diffeomorphism in the plane. In higher
dimensions, one knows that an embedding of the (d − 1) dimensional sphere in a Rd divides
space into two regions. This is the Jordan-Brouwer separation theorem. It is no more true
in general that the two parts are homeomorphic to {x ∈ Rd | |x| < 1}: a counter example
is the Alexander horned sphere which is a topological 2-sphere but where the unbounded
19
component is not simply connected and so not homeomorphic to the complement of a unit ball.
See [56].
49. Chinese remainder theorem

Given integers a, b, a linear modular equation or congruence ax + b = 0 mod m asks to
find an integer x such that ax + b is divisible by m. This linear equation can always be solved
if a and m are coprime. The Chinese remainder theorem deals with the system of linear
modular equations x = b1 mod m1 , x = b2 mod m2 , . . . , x = bn mod mn , where mk are the
moduli. More generally, for an integer n × n matrix A we call Ax = bmod m a Chinese
remainder theorem system or shortly CRT system if the mj are pairwise relatively prime
and in each row there is a matrix element Aij relatively prime to mi .
Theorem: Every Chinese remainder theorem system has a solution.

The classical single variable case case is when Ai1 = 1 and Aij = 0 for j > 1. Let M =
m1 · · · m2 · · · mn be the product. In this one-dimensional case, the result implies that xmod M
→ (x mod m1 , . . . , (x mod mn ) is a ring isomorphism. Define Mi = M/mi . An explicit
algorithm is to finding numbers yi , zi with yi Mi + zi mi = 1 (finding y, z solving ay + bz = 1 for
coprime a, b is computed using the Euclidean algorithm), then finding x = b1 m1 y1 + · · · +
bn mn yn . [116, 289]. The multi-variable version appeared in 2005 [247, 250].
50. Bézout’s theorem

A polynomial is homogeneous if the total degree of all its monomials is the same. A homo-
geneous polynomial f in n + 1 variables of degree d ≥ 1 defines a projective hypersurface
f = 0. Given n projective irreducible hypersurfaces fk = ck of degree dk in a projective space
Pn we can look at the solution set {f = c} = {f1 = c1 , · · · , fk = ck } of a system of nonlinear
equations. The Bézout’s bound is d = d1 · · · dk the product of the degrees. Bézout’s theo-
rem allows to count the number of solutions of the system, where the number of solutions is
counted with multiplicity.
Theorem: The set {f = c} is either infinite or has d elements.
Bézout’s theorem was stated in the “Principia” of Newton in 1687 but proven fist in 1779 by
Étienne Bézout. If the hypersurfaces are all irreducible and in “general position”, then there
are exactly d solutions and each has multiplicity 1. This can be used also for affine surfaces. If
y 2 −x3 −3x−5 = 0 is an elliptic curve for example, then y 2 z −x3 −3xz 2 −5z 3 = is a projective
hypersurface, its projective completion. Bézout’s theorem implies part the fundamental
theorem of algebra as for n = 1, when we have only one homogeneous equation we have d roots
to a polynomial of degree d. The theorem implies for example that the intersection of two conic
sections have in general 2 intersection points. The example x2 − yz = 0, x2 + z 2 − yz = 0
has only the solution x = z = 0, y = 1 but with multiplicity 2. As non-linear systems of
equations appear frequently in computer algebra this theorem gives a lower bound on the
computational complexity for solving such problems.
51. Group theory

A finite group (G, ∗, 1) is a finite set containing a unit 1 ∈ G and a binary operation ∗ :
G × G → G satisfying the associativity property (x ∗ y) ∗ z = x ∗ (y ∗ z) and such that for
20
OLIVER KNILL
every x, there exists a unique y = x−1 such that x ∗ y = y ∗ x = 1. The order n of the group
is the number of elements in the group. An element x ∈ G generates a subgroup formed by
1, x, x2 = x ∗ x, . . . . This is the cyclic subgroup C(x) generated by x. Lagrange’s theorem
tells
Theorem: |C(x)| is a factor of |G|
The origins of group theory go back to Joseph Louis Lagrange, Paulo Ruffini and Évariste
Galois. The concept of abstract group appeared first in the work of Arthur Cayley. Given a
subgroup H of G, the left cosets of H are the equivalence classes of the equivalence relation
x ∼ y if there exists z ∈ H with x = z ∗ y. The equivalence classes G/N partition G.
The number [G : N ] of elements in G/H is called the index of H in G. It follows that
|G| = |H|[G : H] and more generally that if K is a subgroup of H and H is a subgroup of G
then [G : K] = [G : H][H : K]. The group N generated by x is a called a normal group
N / G if for all a ∈ N and all x in G the element x ∗ a ∗ x−1 is in N . This can be rewritten as
H ∗ x = x ∗ H. If N is a normal group, then G/H is again a group, the quotient group. For
example, if f : G → G0 is a group homomorphism, then the kernel of f is a normal subgroup
and |G| = |ker(f )||im(f )| because of the first group isomorphism theorem.
52. Primes
A prime is an integer larger than 1 which is only divisible by 1 or itself. The Wilson theorem
allows to define a prime as a number n for which (n − 1)! + 1 is divisible by n. Euclid already
knew that there are infinitely many primes (if there were finitely many p1 , . . . , pn , the new
number p1 p2 · · · pn + 1 would have a prime factor different
P∞from the given set). It also follows
from the divergence of the harmonic series ζ(1)P= n=1 1/n P = 1 + 1/2 + 1/3 + · · · and
the Euler golden key or Euler product ζ(s) = ∞ n=1 1/n 2
= s −1
p prime (1 − 1/p ) for the
Riemann zeta function ζ(s) that there are infinitely many primes as otherwise, the product
to the right would be finite.
Let π(x) be the prime-counting function which gives the number of primes smaller or equal
to x. Given two functions f (x), g(x) from the integers to the integers, we say f ∼ g, if
limx→∞ f (x)/g(x) = 1. The prime number theorem tells
Theorem: π(x) ∼ x/ log(x).
The result was investigated experimentally first by Anton Ferkel and Jurij Vega, Adrien-Marie
Legendre first conjectured in 1797 a law of this form. Carl Friedrich Gauss wrote in 1849
that he experimented independently around 1792 with such a law. The theorem was proven in
1896 by Jacques Hadamard and Charles de la Vallée Poussin. Proofs without complex analysis
were put forward by Atle Selberg and Paul Erdös in 1949. The prime number theorem also
assures that there are infinitely many primes but it makes the statement quantitative in that
it gives an idea how fast the number of primes grow asymptotically. Under√ the assumption
of the Riemann
Rx hypothesis, Lowell Schoenfeld proved |π(x) − li(x)| < x log(x)/(8π), where
li(x) = 0 dt/ log(t) is the logarithmic integral.
53. Cellular automata

A finite set A called alphabet and an integer d ≥ 1 defines the compact topological space
d
Ω = AZ of all infinite d-dimensional configurations. The topology is the product topology
21
which is compact by the Tychonov theorem. The translation maps Ti (x)n = xn+ei are homeo-
morphisms of Ω called shifts. A closed T invariant subset X ⊂ Ω defines a subshift (X, T ). An
automorphism T of Ω which commutes with the translations Ti is called a cellular automaton,
abbreviated CA. An example of a cellular automaton is a map T xn = φ(xn+u1 , . . . xn+uk ) where
U = {u1 , . . . uk } ⊂ Zd is a fixed finite set. It is called an local automaton because it is defined
by a finite rule so that the status of the cell n at the next step depends only on the status of
the “neighboring cells” {n + u | u ∈ U }. The following result is the Curtis-Hedlund-Lyndon
theorem:
Theorem: Every cellular automaton is a local automaton.
Cellular automata were introduced by John von Neumann and mathematically in 1969 by
Hedlund [193]. The result appears there. Hedlund saw cellular automata also as maps on
subshifts. One can so look at cellular automata on subclasses of subshifts. For example,
one can restrict the cellular automata map T on almost periodic configurations, which are
subsets X of Ω on which (X, T1 , . . . Tj ) has only invariant measures µ for which the Koopman
operators Ui f = f (Ti ) on L2 (X, µ) have pure point spectrum. A particularly well studied case
is d = 1 and A = {0, 1}, if U = {−1, 0, 1}, where the automaton is called an elementary
cellular automaton. The Wolfram numbering labels the 28 possible elementary automata
with a number between 1 and 255. The game of life of Conway is a case for d = 2 and
A = {−1, 0, 1} × {−1, 0, 1}. For literature on cellular automata see [432] or as part of complex
systems [433] or evolutionary dynamics [318]. For topological dynamics, see [109].
54. Topos theory

A category has objects as nodes and morphisms as arrows going from one object to an
other object. There can be multiple connections and self-loops so that one can visualize a
category as a quiver. Every object has the identity arrow 1A . A topos X is a Cartesian
closed category C in which finite limits exists and which has a sub-object classifier Ω
allowing to identify sub-objects with morphisms from X to Ω. Cartesian closed means
that one can define for any pair of objects A, B in C the product A × B and an equalizer
representing solutions f = g to arrows f : A → B, G : A → B as well as an exponential
B A representing all arrows from A to B. An example is the topos of sets. An example of a
sub-object classifier is Ω = {0, 1} encoding “true or false”.
The slice category E/X of a category E with an object X in E is a category, where the objects
are the arrows from E → X. An E/X arrow between objects f : A → X and g : B → X is
a map s : A → B which produces a commutative triangle in E. The composition is pasting
triangles together. The fundamental theorem of topos theory is:
Theorem: The slice category E/X of a topos E is a topos.
For example, if E is the topos of sets, then the slice category is the category of pointed
sets: the objects are then sets together with a function selecting a point as a “base point”.
A morphism f : A → B defines a functor E/B → E/A which preserves exponentials and the
subobject classifier Ω. Topos theory was motivated by geometry (Grothendieck), physics
(Lawvere), topology (Tierney) and algebra (Kan). It can be seen as a generalization and
even a replacement of set theory: the Lawvere’s elementary theory of the category of
sets ETCS is seen as part of ZFC which are less likely to be inconsistent [276]. For a short
introduction [214], for textbooks [295, 73], for history of topos theory in particular, see [294].
22
OLIVER KNILL
55. Transcendentals
A root of an equation f (x) = 0 with integer polynomial f (x) = an xn + an−1 xn−1 + · · · + a0
with n ≥ 0 and aj ∈ Z is called an algebraic number. The set A of algebraic numbers is
sub-field of the field R of real numbers. The field A s the algebraic closure of the rational
numbers Q. It is of number theoretic interest as it contains all algebraic number fields, finite
degree field extensions of Q. The complement R \ A is the set of transcendental numbers.
Transcendental numbers are necessarily irrational because every rational number x = p/q is
algebraic, solving qx − p = 0. Because the set of algebraic numbers is countable and the real
numbers are not, most numbers are transcendental. The group of all automorphisms of A which
fix Q is called the absolute Galois group of Q.
Theorem: π and e are transcendental
This result is due to Ferdinand von Lindemann. He proved that ex is transcendental for every
non-zero algebraic number x. This immediately implies e is transcendental. Now, if π were
algebraic, then πi would be algebraic and eiπ = −1 would be transcendental. But −1 is
rational. Lindemann’s result was extended in 1885 by Karl Weierstrass to the statement telling
that if x1 , . . . xn are linearly independent algebraic numbers, then ex1 , . . . exn are algebraically
independent. The transcendental property of π also proves that π is irrational. This is easier
to prove directly. See [204].
56. Recurrence
A homeomorphism T : X → X of a compact topological space X defines a topological
dynamical system (X, T ). We write T j (x) = T (T (. . . T (x))) to indicate that the map T is
applied j times. For any d > 0, we get from this a set (T1 , T2 , . . . , Td ) of commuting homeo-
morphisms on X, where Tj (x) = T j x. A point x ∈ X is called multiple recurrent for T if for
every d > 0, there exists a sequence n1 < n2 < n3 < · · · of integers nk ∈ N for which Tjnk x → x
for k → ∞ and all j = 1, . . . , d. Fürstenberg’s multiple recurrence theorem states:
Theorem: Every topological dynamical system is multiple recurrent.
It is known even that the set of multiple recurrent points are Baire generic. Hillel Fürstenberg
proved this result in 1975. There is a parallel theorem for measure preserving systems:
an automorphism T of a probability space (Ω, A, P) is called multiple recurrent if there
exists A ∈ A and an integer n such that P[A ∩ T1 (A) ∩ · · · ∩ Td (A)] > 0. This generalizes the
Poincaré recurrence theorem, which is the case d = 1. Recurrence theorems are related
to the Szemerédi theorem telling that a subset A of N of positive upper density contains
arithmetic progressions of arbitrary finite length. See [154].
57. Solvability
A basic task in mathematics is to solve polynomial equations p(x) = an xn + an−1 xn−1 +
· · · + a1 x + a0 = 0 with complex coefficients ak using explicit formulas involving roots. One
calls this an explicit algebraic solution. The √ linear case ax + b = 0 with x = −b/a, the
2
quadratic case ax + bx + c = 0 with x = (−b ± b2 − 4ac)/(2a) were known since antiquity.
The cubic x3 + ax2 + bx + C = 0 was solved by Niccolo Tartaglia and Cerolamo Cardano: a
first substitution x = X − a/3 produces the depressed cubic X 3 + pX + q (first solved by
Scipione dal Ferro). The substitution X = u − p/(3u) then produces a quadratic equation for
23
u3 . Lodovico Ferrari solved finally the quartic by reducing it to the cubic. It was Paolo Ruffini,
Niels Abel and Évariste Galois who realized that there are no algebraic solution formulas any
more for polynomials of degree n ≥ 5.
Theorem: Explicit algebraic solutions to p(x) = 0 exist if and only if n ≤ 4.
The quadratic case was settled over a longer period in independent development in Babylonian,
Egyptian, Chinese and Indian mathematics. The cubic and quartic discoveries were dramatic
culminating with Cardano’s book of 1545, marking the beginning of modern algebra. After
centuries of failures of solving the quintic, Paolo Ruffini published the first proof in 1799, a
proof which had a gap but who paved the way for Niels Hendrik Abel and Évariste Galois. For
further discoveries see [290, 280, 10].
58. Galois theory

If F is sub-field of E, then E is a vector space over F . The dimension of this vector space is
called the degree [E : F ] of the field extension E/F . The field extension is called finite
if [E : F ] is finite. A field extension is called transcendental if there exists an element
in E which is not a root of an integral polynomial f with coefficients in F . Otherwise, the
extension is called algebraic. In the later case, there exists a unique monique polynomial f
which is irreducible over F and the field extension is finite. An algebraic field extension E/F
is called normal if every irreducible polynomial over K with at least one root in E splits over
F into linear factors. An algebraic field extension E/F is called separable if the associated
0
irreducible polynomial f is separable, meaning thatP f ispk not zero. This means, that F has
zero characteristic or that f is not of the form k ak x if F has characteristic p. A field
extension is called Galois if it normal and separable. Let Fields(E/F ) be the set of subfields
of E/F and Groups(E/F )) the set of subgroups of the automorphism group Aut(E/F ). The
Fundamental theorem of Galois theory assures:
bijective
Theorem: Fields(E/F ) ↔ Groups(E/F ) if E/F is Galois.
The intermediate fields of E/F are so described by groups. It implies the Abel-Ruffini the-
orem about the non-solvability of the quintic by radicals. The fundamental theorem demon-
strates that solvable extensions correspond to solvable groups. The symmetry groups of
permutations of 5 or more elements are no more solvable. See [389].
59. Metric spaces

A topological space (X, O) is given by a set X and a finite collection O of subsets of X with
the property that the empty set ∅ and Ω both belong to O and that O is closed under arbitrary
unions and finite intersections. The sets in O are called open sets. Metric spaces (X, d) are
special topological spaces. In that case, O consists of all sets U such that for every x ∈ U there
exists r > 0 such that the open ball Br (x) = {y ∈ X | d(x, y) < r} is contained in U . Two
topological spaces (X, O), (Y, Q) are homeomorphic if there exists a bijection f : X → Y ,
such that f and f −1 are both continuous. A function f : X → Y is continuous if f −1 (A) ∈ O
for all A ∈ Q. When is a topological space homeomorphic to a metric space? The Urysohn
metrization theorem gives an answer: we need the regular Hausdorff property meaning
that a closed set K and a point x can be separated by disjoint neighborhoods K ⊂ U, y ∈ V .
We also need the space to be second countable meaning that there is a countable base (a
24
OLIVER KNILL
base in O is a subset B ⊂ O such that every U ∈ O can be written as a union of elements in

B.)
Theorem: A second countable regular Hausdorff space is metrizable.

The result was proven by Pavel Urysohn in 1925 with “regular” replaced by “normal” and by
Andrey Tychonov in 1926. It follows that a compact Hausdorff space is metrizable if and only
if it is second countable. For literature, see [56].
60. Fixed point

Given a continuous transformation T : X → X of a compact topological space X, one
can look for the fixed point set FixT (X) = {x | T (x) = x}. This is useful for finding
periodic points as fixed points of T n = T ◦ T ◦ T · · · ◦ T are periodic points of period n.
If X has a finite cohomology like if X is a compact d-manifold with boundary, one can
look at the linear map Tp induced on the cohomology groups H p (X). The super trace
χT (X) = dp=0 (−1)p tr(Tp ) is called the Lefschetz number of T on X. If T is the identity,
P
this is the Euler characteristic. Let indT (x) be the Brouwer degree of the map T induced
on a small (d − 1)-sphere S around x. This is the trace of the linear map Td−1 induced from
T on the cohomology group H d−1 (S) which is an integer. If T is differentiable and dT (x) is
invertible, the Brouwer degree is indT (x) = sign(det(dT )). Let FixT (X) denote the set of fixed
points of T . The Lefschetz-Hopf fixed point theorem is
P
Theorem: If FixT (X) is finite, then χT (X) = x∈FixT (X) indT (x).
A special case is the Brouwer fixed point theorem: if X is a compact convex subset of
Euclidean space. In that case χT (X) = 1 and the theorem assures the existence of a fixed
point. In particular, if T : D → D is a continuous map from the disc D = {x2 + y 2 ≤ 1}
onto itself, then T has a fixed point. The Brouwer fixed point theorem was proved in
1910 by Jacques Hadamard and Luitzen Egbertus Jan Brouwer. The Schauder fixed point
theorem from 1930 generalizes the result to convex compact subsets of Banach spaces. The
Lefschetz-Hopf fixed point theorem was given in 1926. For literature, see [118, 48].
61. Quadratic reciprocity

Given a prime p, a number a is called a quadratic residue if there exists a number x such
that x2 has remainder a modulo p. In other words quadratic residues are the squares in the
field Zp . The Legendre symbol (a|p) is defined by be 0 if a is 0 or a multiple of p and 1 if
a is a non-zero residue of p and −1 if it is not. While the integer 0 is sometimes considered
to be a quadratic residue we don’t include it as it is a special case. Also, in the multiplicative
group Z∗p without zero, there is a symmetry: there are the same number of quadratic residues
and non-residues. This is made more precise in the law of quadratic reciprocity
p−1 q−1
Theorem: For any two odd primes (p|q)(q|p) = (−1) 2 2 .
This means that (p|q) = −(q|p) if and only if both p and q have remainder 3 modulo 4. The
odd primes with of the form 4k + 3 are also prime in the Gaussian integers. To remember
the law, one can think of them as “Fermions” and quadratic reciprocity tells they Fermions
are anti-commuting. The odd primes of the form 4k + 1 factor by the 4-square theorem
25
in the Gaussian plane to p = (a + ib)(a − ib) and are as a product of two Gaussian primes
and are therefore Bosons. One can remember the rule because Boson commute both other
particles so that if either p or q or both are “Bosonic”, then (p|q) = (q|p). The law of quadratic
reciprocity was first conjectured by Euler and Legendre and published by Carl Friedrich Gauss
in his Disquisitiones Arithmeticae of 1801. (Gauss found the first proof in 1796). [187, 204].
62. Quadratic map

Every quadratic map z → f (z) = z 2 + bz + d in the complex plane is conjugated to one of
the quadratic family maps Tc (z) = z 2 + c. The Mandelbrot set M = {c ∈ C, Tcn (0) stays
bounded } is also called the connectedness locus of the quadratic family because for c ∈ M ,
the Julia set Jc = {z ∈ C; T n (z) stays bounded } is connected and for c ∈
/ M , the Julia set Jc
is a Cantor set. The fundamental theorem for quadratic dynamical systems is:
Theorem: The Mandelbrot set is connected.
Mandelbrot first thought after seeing experiments that it was disconnected. The theorem is due
to Adrien Duady and John Hubbard in 1982. One can also look at the connectedness locus for
T (z) = z d + c, which leads to Multibrot sets or the map z → z + c, which leads to the tricorn
or mandelbar which is not path connected. One does not know whether the Mandelbrot set
M is locally connected, nor whether it is path connected. See [300, 74, 34]
63. Differential equations

Let us say that a differential equation x0 (t) = F (x) is integrable if a trajectory x(t) either
converges to infinity, or to an equilibrium point or to a limit cycle or limiting torus,
where it is a periodic or almost periodic trajectory. We assume F has global solutions. The
Poincaré-Bendixon theorem is:
Theorem: Any differential equation in the plane is integrable.
This changes in dimensions 3 and higher. The Lorenz attractor or the Rössler attractor are
examples of strange attractors, limit sets on which the dynamics can have positive topological
entropy and is therefore no more integrable. The theorem also does not hold any more on two
dimensional tori as there can be recurrent non-periodic orbits and even weak mixing. The proof
of the Poincaré-Bendixon theorem relies on the Jordan curve theorem. [88, 230].
64. Approximation theory

A function f on a closed interval I = [a, b] is called continuous if for every > 0 there exists
a δ > 0 such that if |x − y| < δ then |f (x) − f (y)| < . In the space X = C(I) of all continuous
functions, one can define a distance d(f, g) = maxx∈I |f (x) − g(x)|. A subset Y of X is called
dense if for every > 0 and every x ∈ X, there exists y ∈ Y with d(x, y) < . Let P denote
the class of polynomials in X. The Weierstrass approximation theorem tells that
Theorem: Polynomials P are dense in continuous functions C(I).
The Weierstrass theorem has been proven in 1885 by Karl Weierstrass. A constructive proof sug-
gested by sergey Bernstein in 1912 uses Bernstein polynomials fn (x) = nk=0 f (k/n)Bk,n (x)
P
with Bk,n (x) = B(n, k)xk (1 − x)n−k , where B(n, k) denote the Binomial coefficients. The result
has been generalized to compact Hausdorff spaces X and more general subalgebras of C(X).
26
OLIVER KNILL
The Stone-Weierstrass approximation theorem was proven by Marshall Stone in 1937

and simplified in 1948 by Stone. In the complex, there is Runge’s theorem from 1885 ap-
proximating functions holmomorphic on a bounded region G with rational functions uniformly
on a compact subset K of G and Mergelyan’s theorem from 1951 allowing approximation
uniformly on a compact subset with polynomials if the region G is simply connected. In nu-
merical analysis one has the task to approximate a given function space by functions from a
simpler class. Examples are approximations of smooth functions by polynomials, trigonometric
polynomials. There is also the interpolation problem of approximating a given data set with
polynomials or piecewise polynomials like splines or Bézier curves. See [405, 313].
65. Diophantine approximation

An algebraic number is a root of a polynomial p(x) = an xn + an−1 xn−1 + · · · + a1 x + a0 with
integer coefficients ak . A real number x is called Diophantine if there exists > 0 and a
positive constant C such that the Diophantine condition |x − p/q| > C/q 2+ is satisfied for
all p, and all q > 0. Thue-Siegel-Roth theorem tells:
Theorem: Any irrational algebraic number is Diophantine.
The Hurwitz’s theorem
√ from 1891 assures that there are infinitely many p, q with |x − p/q| <
2
C/q for C = 1/ 5. This shows that the Tue-Siegel-Roth √ Theorem can not be extended to
= 0. The√Hurwitz constant C is optimal. For any C < 1/ 5 one can with the golden ratio
x = (1 + 5)/2 have only finitely many p, q with |x − p/q| < C/q 2 . The set of Diophantine
numbers has full Lebesgue measure. A slightly larger set isPthe Brjuno set of all numbers
for which the continued fraction convergent pn /qn satisfies n log(qn+1 )/qn < ∞. A Brjuno
rotation number assures the Siegel linearization theorem still can be proven. For quadratic
polynomials, Jean-Christophe Yoccoz showed that linearizability implies the rotation number
must be Brjuno. [74, 196]
66. Almost periodicity

R
If µ is a probability measure of compact support on R, then µ̂n = einx dµ(x) are the
Fourier coefficients of µ. The Riemann-Lebesgue lemma tells that if µ is absolutely
continuous, then µ̂n goes to zero. The pure point part can be detected with the following
Wiener theorem:
Theorem: limn→∞ n1 nk=1 |µ̂k |2 = x∈T |µ({x})|2 .
P P
P P ˆ ˆ
This looks a bit like the Poisson summationP formula P n f (n) = n f (n), where f is the
Fourier transform of f . [The later follows from n e2πikx = n δ(x − n), where δ(x) is a Dirac
delta function. The Poisson formula holds if f is uniformly continuous and if both f and fˆ
satisfy the growth condition |f (x)| ≤ C/|1 + |x||1+ . ] More generally, one can read off the
Hausdorff dimension from decay rates of the Fourier coefficients. See [236, 387].
67. Shadowing
Let T be a diffeomorphism on a smooth Riemannian manifold M with geodesic metric
d. A T -invariant set is called hyperbolic if for each x ∈ K, the tangent space Tx M splits
into a stable and unstable bundle Ex+ ⊕ Ex− such that for some 0 < λ < 1 and constant C,
one has dT Ex± = ET±x and |dT ±n v| ≤ Cλn for v ∈ E ± and n ≥ 0. An -orbit is a sequence
27
xn of points in M such that xn+1 ∈ B (T (xn )), where B is the geodesic ball of radius . Two
sequences xn , yn ∈ M are called δ-close if d(yn , xn ) ≤ δ for all n. We say that a set K has the
shadowing property, if there exists an open neighborhood U of K such that for all δ > 0
there exists > 0 such that every -pseudo orbit of T in U is δ-close to true orbit of T .
Theorem: Every hyperbolic set has the shadowing property.
This is only interesting for infinite K as if K is a finite periodic hyperbolic orbit, then the orbit
itself is the orbit. It is interesting however for a hyperbolic invariant set like a Smale horse
shoe or in the Anosov case, when the entire manifold is hyperbolic. See [230].
68. Partition function

Let p(n) denote the number of ways we can write n as a sum of positivePintegers with-
out
Q∞ distinguishing the order. Euler used its generating function which is ∞ n
n=0 p(n)x =
k −1 2 3
k=1 (1 − x ) . The reciprocal function (1 − x)(1 − x ) + (1 − x ) · · · is Pcalled the Euler func-
tion and generates the generalized Pentagonal number theorem k∈Z (−1)k xk(3k−1)/2 =
1 − x − x2 + x5 − x7 − x12 − x15 · · · leading to the recursion p(n) = p(n − 1) + p(n − 2) − p(n −
5) − p(n − 7) + p(n − 12) + p(n − 15) · · · . The Jacobi triple product identity is
Q∞ 2m
)(1 − x2m−1 y 2 )(1 − x2m−1 y −2 ) = ∞ n2 2n
P
Theorem: n=1 (1 − x n=−∞ x y .
√ √
The formula was found in 1829 by Jacobi. For x = z z and y 2 = − z the identity reduces to
the pentagonal number theorem. See [16].
69. Burnside lemma

If G is a finite group acting on a finite set X, let X/G denote the number of disjoint orbits
and X g = {x ∈ X | g.x = x, ∀g ∈ G} the fixed point set of elements which are fixed by g.
The number |X/G| of orbits and the group order |G| and the size of the fixed point sets
are related by the Burnside lemma:
1 g
P
Theorem: |X/G| = |G| g∈G |X |
The result was first proven by Frobenius in 1887. Burnside popularized it in 1897 [69].
70. Taylor series

A complex-valued function f which is analytic in a disc D = Dr (a) = {|x − a| < r} can be
written as a series involving the n’th derivatives f (n) (a) of f at a. If f is real valued on the real
axes, the function is called real analytic in (x − a, x + a). In several dimensions we can use
multi-index notation a = (a1 , . . . , ad ), n = (n1 , . . . , nd ), x = (x1 , . . . , xd ) and xn = xn1 1 · · · xnd d
and f (n) (x) = ∂xn11 · · · ∂xndd and use a polydisc D = Dr (a) = {|x1 − a1 | < r1 , . . . |xd − ad | < rd }.
The Taylor series formula is:
P∞ f (n) (a)
Theorem: For analytic f in D, f (x) = n=0 n!
(x − a)n .
Here, Tr (a) = {|xi − a1 | = r1 . . . |xd − ad | = rd } isPthe boundary torus. For example, for
f (x) = exp(x), where f (n) (0) = 1, one has f (x) = ∞ n
n=0 x /n!. Using the differential op-
(n)
∞ f (x) n
erator Df (x) = f 0 (x), one can see f (x + t) = t = eDt f (x) as a solution of
P
n=0 n!
28
OLIVER KNILL
the transport equation R ft = Df . One can also represent f as a Cauchy formula for
polydiscs 1/(2πi)d |Tr (a)| f (z)/(z − a)d dz integrating along the boundary torus. Finite Taylor
series hold in the case if f is m + 1 times differentiable. In that case one has a finite series
f (n) (a)
S(x) = m (x − a)n such that the Lagrange rest term is f (x) − S(x) = R(x) =
P
n=0 n!
f m+1 (ξ)(x − a)m+1 /((m + 1)!), where ξ is between x and a. This generalizes the mean value
theorem inR the case m = 0, where f is only differentiable. The remainder term can also be
x
written as a f (m+1) (s)(x − a)m /m! ds. Taylor did state but not justify the formula in 1715
which was actually a difference formula. 1742 Maclaurin uses the modern form. [262].
71. Isoperimetric inequality

Given a smooth surface S in Rn homeomorphic to a sphere and bounding a region B. Assume
that the surface area |S| is fixed. How large can the volume |B| of B become? If B is the
unit ball B1 with volume |B1 | the answer is given by the isoperimetric inequality:
Theorem: nn |B|n−1 ≤ |S|n /|B1 |.
If B = B1 , this gives n|B| ≤ |S|, which is an equality as then the volume of the ball |B| =
π n/2 /Γ(n/2+1) and the surface area of the sphere |S| = nπ n/2 /Γ(n/2+1) which Archimedes
first got in the case n = 3, where |S| = 4π and |B| = 4π/3. The classical isoperimetric
problem is n = 2, where we are in the plane R2 . The inequality tells then 4|B| ≤ |S|2 /π
which means 4πArea ≤ Length2 . The ball B1 with area 1 maximizes the functional. For
n = 3, with usual Euclidean space R3 , the inequality tells |B|2 ≤ (4π)3 /(27 · 4π/3) which is
|B| ≤ 4π/3. The first proof in the case n = 2 was attempted by Jakob Steiner in 1838 using
the Steiner symmetrization process which is a refinement of the Archimedes-Cavalieri
principle. In 1902 a proof by Hurwitz was given using Fourier series. The result has been
extended to geometric measure theory [144]. One can also look at the discrete problem to
maximize the area defined by a polygon: if {(xi , yi ), i = P0, . . . n − 1} are the points of the
polygon, then the area is given by Green’s formula as A = n−1 i=0 xi yi+1 − xi+1 yi and the length
Pn−1 2 2
is L = i=0 (xi − xi+1 ) + (yi − yi+1 ) with (xn , yn ) identified with (x0 , y0 ). The Lagrange
equations for A under the constraint L = 1 together with a fix of (x0 , y0 ) and (x1 = 1/n, 0)
produces two maxima which are both regular polygons. A generalization to n-dimensional
Riemannian manifolds is given by the Lévi-Gromov isoperimetric inequality.
72. Riemann Roch

A Riemann surface is a one-dimensional complex manifold. It is a two-dimensional real analytic
manifold but it has also a complex structure forcing it to be orientable for example. Let
G be a compact connected Riemann surface of Euler characteristic χ(G) = 1 − g, where
g = b1 (G) is the genus, the number of handles P of G (and 1 = b0 (G) indicates that we have
only one connected component). A divisor D = i ai zi on G is an element of the free Abelian
P ai ∈ Z
group on the points of the surface. These are finite formal sums of points zi in G, where
is the multiplicity of the point zi . The degree of the divisor is defined as deg(D) = i ai . Let
us write χ(X) = deg(D) + χ(G) = deg(D) + 1 − g and call this the Euler characteristic of
the divisor D as one can see a divisor as a geometric object by itself generalizing the complex
manifold X (which is thePcase D P = 0). A meromorphic function f on G defines the
principal divisor (f ) = i ai zi − j bj wj , where ai are the multiplicities of the roots zi of
f and bj the multiplicities of the poles wj of f . The principal divisor of a global meromorphic
29
1-form dz which is called the canonical divisor K. Let l(D) be the dimension of the linear
space of meromorphic functions f on G for which (f ) + D ≥ 0. (The notation ≥ 0 means that
all coefficients are non-negative. One calls such a divisor effective). The Riemann-Roch
theorem is
Theorem: l(D) − l(K − D) = χ(D)
The idea of a Riemann surfaces was defined by Bernhard Riemann. Riemann-Roch was proven
for Riemann surfaces by Bernhard Riemann in 1857 and Gustav Roch in 1865. It is possible
to see this as a Euler-Poincaré type relation by identifying the left hand side as a signed
cohomological Euler characteristic and the right hand side as a combinatorial Euler character-
istic. There are various generalizations, to arithmetic geometry or to higher dimensions. See
[172, 358].
73. Optimal transport
Given two probability spaces (X, P ), (Y, Q) and a continuous cost function c : X ×Y → [0, ∞],
the optimal transport
R problem or Monge-Kantorovich minimization problem is to find
the minimum of X c(x, T (x)) dP (x) among all coupling transformations T : X → Y which
have the property that it transports the measure P to the measure Q. More generally, one
looks at a measure π on X × Y such that the projection of π Ronto X it is P and the projection
of π onto Y is Q. The function to optimize is then I(π) = X×Y c(x, y) dπ(x, y). One of the
fundamental results is that optimal transport exists. The technical assumption is that if the
two probability spaces X, Y are Polish (=separable complete metric spaces) and that the cost
function c is continuous.
Theorem: For continuous cost functions c, there exists a minimum of I.
In the simple set-up of probability spaces, this just follows from the compactness (Alaoglu
theorem for balls in the weak star topology of a Banach space) of the set of probability measures:
any sequence πn of probability measures on X × Y has a convergent subsequence. Since I
is continuous, picking a sequence πn with I(πn ) decreasing produces to a minimum. The
problem was formalized in 1781 by Gaspard Monge and worked on by Leonid Kantorovich.
Hirisho Tanaka in the 1970ies produced connections with partial differential equations like the
Bolzmann equation. There are also connections to weak KAM theory in the form of Aubry-
Mather theory. The above existence result is true under substantial less regularity. The question
of uniqueness or the existence of a Monge coupling given in the form of a transformation T is
subtle [412].
74. Structure from motion
Given m hyper planes in Rd serving as retinas or photographic plates for affine cameras and
n points in Rd . The affine structure from motion problem is to understand under which
conditions it is possible to recover both the points and planes when knowing the orthogonal
projections onto the planes. It is a model problem for the task to reconstruct both the scene
as well as the camera positions if the scene has n points and m camera pictures were taken.
Ullman’s theorem is a prototype result with n = 3 different cameras and m = 3 points which are
not collinear. Other setups are perspective cameras or omni-directional cameras. The
Ullman map F is a nonlinear map from Rd·2 × SOd2 to (R3d−3 )2 which is a map between equal
dimensional spaces if d = 2 and d = 3. The group SOd is the rotation group in R describing the
30
OLIVER KNILL
possible ways in which the affine camera can be positioned. Affine cameras capture the same
picture when translated so that the planes can all go through the origin. In the case d = 2, we
get a map from R4 × SO22 to R6 and in the case d = 3, F maps R6 × SO32 into R12 .
Theorem: The structure from motion map is locally invertible.
In the case d = 2, there is a reflection ambiguity. In dimension d = 3, the number of ambiguities

is typically 64. Ullman’s theorem appeared in 1979 in [408]. Ullman states the theorem for d=3
with 4 points as adding a four point cuts the number of ambiguities from 64 to 2. See [255]
both in dimension d=2 and d=3 the Jacobean dF of the Ullman map is seen to be invertible
and the inverse of F is given explicitly. For structure from motion problems in computer vision
in general, see [143, 188, 406]. In applications one takes n and m large and reconstructs both
the points as well as the camera parameters using statistical data fitting.
75. Poisson equation

What functions u solve the Poisson equation −∆u = f , a Rpartial differential equation? The
right hand side can be written down for f ∈ L1 as Kf (x) = Rn G(x, y)f (y) dy + h, where h is
harmonic. If f = 0, then the Poisson equation is the Laplace equation. The function G(x, y)
is the Green’s function, an integral kernel. It satisfies −∆G(x, y) = δ(y −x), where δ is the
Dirac delta function, a distribution. It is given by G(x, y) = − log |x − y|/(2π) for n = 2 or
G(x, y) = |x − y|−1 /(4π) for n = 3. In elliptic regularity theory, one replaces the Laplacian
−∆ with an elliptic second order differential operator L = A(x) · D · D + b(x) · D + V (x)
where D = ∇ is the gradient and A is a positive definite matrix, b(x) is a vector field and c is
a scalar field.
Theorem: For f ∈ Lp and p > n, then Kf is differentiable.
The result is much more general and can be extended. If f is in C k and has compact support
for example, then Kf is in C k+1 . An example of the more general set up is the Schrödinger
operator L = −∆ + V (x) − E. The solution to Lu = 0, solves then an eigenvalue problem.
As one looks for solutions in L2 , the solution only exists if E is an eigenvalue of L. The
Euclidean space Rn can be replaced by a bounded domain Ω of Rn where one can look at
boundary conditions like of Dirichlet or von Neumann type. Or one can look at the situation
on a general Riemannian manifold M with orRwithout boundary. On a Hilbert space, one has
then Fredholm theory. The equation P un = nG(x, y)f (y)dy is called a Fredholm integral
equation and det(1 − sG) = exp(− n s tr(G )/n!) the Fredholm determinant leading to
the zeta function 1/det(1 − sG). See [340, 279].
76. Four square theorem

Waring’s problem asked whether there exists for every k an integer g(k) such that every
positive integer can be written as a sum of g(k) powers xk1 + · · · + xkg(k) . Obviously g(1) = 1.
David Hilbert proved in 1909, that g(k) is finite. This is the Hilbert-Waring theorem. The
following theorem of Lagrange tells that g(2) = 4:
Theorem: Every positive integer is a sum of four squares
31
.
The result needs only to be verified for prime numbers as N (a, b, c, d) = a2 + b2 + c2 + d2
is a norm for quaternions q = (a, b, c, d) which has the property N (pq) = N (p)N (q). This
property can be seen also as a Cauchy-Binet formula, when writing quaternions as complex
2 × 2 matrices. The four-square theorem had been conjectured already by Diophantus, but
was proven first by Lagrange in 1770. The case g(3) = 9 was done by Wieferich in 1912. It is
conjectured that g(k) = 2k + [(3/2)k ] − 2, where [x] is the integral part of a real number. See
[111, 112, 204].
77. Knots
A knot is a closed curve in R3 , an embedding of the circle in three dimensional Euclidean
space. One also draws knots in the 3-sphere S 3 . As the knot complement S 3 − K of a knot
K characterizes the knot up to mirror reflection, the theory of knots is part of 3-manifold
theory. The HOMFLYPT polynomial P of a knot or link K is defined recursively using
skein relations lP (L+ ) + l−1 P (L− ) + mP (L0 ) = 0. Let K#L denote the knot sum which
is a connected sum. Oriented knots form with this operation a commutative monoid with
unknot as unit. It features a unique prime factorization. The unknot has P (K) = 1, the
unlink has P (K) = 0. The trefoil knot has P (K) = 2l2 − l4 + l2 m2 .
Theorem: P (K#L) = P (K)P (L).
The Alexander polynomial was discovered in 1928 and initiated classical knot theory. John
Conway showed in the 60ies how to compute the Alexander polynomial using a recursive skein
relations (skein comes from French escaigne=hank of yarn). The Alexander polynomial allows
to compute an invariant for knots by looking at the projection. The Jones polynomial found
by Vaughan Jones came in 1984. This is generalized by the HOMFLYPT polynomial named
after Jim Hoste, Adrian Ocneanu, Kenneth Millett, Peter J. Freyd and W.B.R. Lickorish from
1985 and J. Przytycki and P. Traczyk from 1987. See [5]. Further invariants are Vassiliev
invariants of 1990 and Kontsevich invariants of 1993.
78. Hamiltonian dynamics

Given a probability space (M, A, m) and a smooth Lie manifold N with potential function
V : N → R, the Vlasov R Hamiltonian differential equations on all maps X = (f, g) : M →
T ∗ N is f 0 = g, g 0 = N ∇V (f (x) − f (y)) dm(y). Starting with X0 = Id, we get a flow Xt
and by push forward an evolution P t = Xt∗ m of probability measures on N . The Vlasov intro-
differential equations on measures in T ∗ N are Ṗ t (x, y) + y · ∇x P t (x, y) − W (x) · ∇y P t (x, y) = 0
with W (x) = M ∇x V (x − x0 )P t (x0 , y 0 )) dy 0 dx0 . Note that while Xt is an infinite dimensional
R
ordinary differential equations evolving maps M → T ∗ N , the path P t is an integro

differential equation describing the evolution of measures on T ∗ N .
Theorem: If Xt solves the Vlasov Hamiltonian, then P t = Xt∗ m solves Vlasov.
This is a result which goes back to James Clerk Maxwell. Vlasov dynamics was introduced in
1938 by Anatoly Vlasov. An existence result was proven by W. Brown and Klaus Hepp in 1977.
The maps Xt will stay perfectly smooth if smooth initially. However, even if P 0 is smooth,
the measure P t in general rather quickly develops singularities so that the partial differential
equation has only weak solutions. The analysis of P directly would involve complicated
32
OLIVER KNILL
function spaces. The fundamental theorem of Vlasov dynamics therefore plays the role
of the method of characteristics in this field. If M is a finite probability space, then the
Vlasov Hamiltonian system is the Hamiltonian n-body problem on N . An other example is
M = T ∗ N and where m is an initial phase space measure. Now Xt is a one parameter family of
diffeomorphisms Xt : M → T ∗ N pushing forward m to a measure P t on the cotangent bundle.
If M is a circle then X 0 defines a closed curve on T ∗ N . In particular, if γ(t) is a curve in N and
X 0 (t) = (γ(t), 0), we have a continuum of particles initially at rest which evolve by interacting
with a force ∇V . About interacting particle dynamics, see [380].
79. Hypercomplexity
A hypercomplex algebra is a finite dimensional algebra over R which is unital and dis-
tributive. The classification of hypercomplex algebras (up to isomorphism) of two-dimensional
hypercomplex algebras over the reals are the complex numbers x + iy with i2 = −1, the
split complex numbers x + jy with j 2 = −1 and the dual numbers (the exterior algebra)
x + y with 2 = 0. A division algebra over a field F is an algebra over F in which division
is possible. Wedderburn’s little theorem tells that a finite division algebra must be a finite
field. Only C is the only two dimensional division algebra over R. The following theorem of
Frobenius classifies the class X of finite dimensional associative division algebras over R:
Theorem: X consists of the algebras R, C and H.
Hypercomplex numbers like quaternions, tessarines or octonions extend the algebra of

complex numbers. Cataloging them started with Benjamin Peirce 1872 ”Linear associative
algebra”. Dual numbers were introduced in 1873 by William Clifford. The Cayley-Dickson
constructions generates iteratively algebras of twice the dimensions: like the complex numbers
from the reals, the quaternions from the complex numbers or the octonions from the quaternions
(for octonions associativity is lost). The next step leads to sedenions but the later are not
even an alternative algebra any more. The Hurwitz and Frobenius theorems limit the number
in the case of normed division algebras. Ferdinand George Frobenius classified in 1877 the
finite-dimensional associative division algebras. Adolf Hurwitz proved in 1923 (posthumously)
that unital finite dimensional real algebra endowed with a positive-definite quadratic form
(a normed division algebra must be R, C, H or O). These four are the only Euclidean
Hurwitz algebras. In 1907, Joseph Wedderburn classified simple algebras (simple meaning
that there are no non-trivial two-sided ideals and ab = 0 implies a = 0 or b = 0). In 1958 J.
Frank Adams showed topologically that R, C, H, O are the only finite dimensional real division
algebras. In general, division algebras have dimension 1, 2, 4 or 8 as Michel Kervaire and Raoul
Bott and John Milnor have shown in 1958 by relating the problem to the parallelizability of
spheres. The problem of classification of division algebras over a field F led Richard Brauer
to the Brauer group BR(F ), which Jean Pierre Serre identified it with Galois cohomology
H 2 (K, K ∗ ), where K ∗ is the multiplicative group of K seen as an algebraic group. Each Brauer
equivalence class among central simple algebras (Brauer algebras) contains a unique division
algebra by the Artin-Wedderburn theorem. Examples: the Brauer group of an algebraically
closed field or finite field is trivial, the Brauer group of R is Z2 . Brauer groups were later
defined for commutative rings by Maurice Auslander and Oscar Goldman and by Alexander
Grothendieck in 1968 for schemes. Ofer Gabber extended the Serre result to schemes with
ample line bundles. The finiteness of the Brauer group of a proper integral scheme is open. See
[27, 141].
33
80. Approximation
The Kolmogorov-Arnold superposition theorem shows that continuous functions C(Rn )
of several variables can be written as a composition of continuous functions of two variables:
Theorem: Every f ∈ C(Rn ) composition of continuous functions in C(R2 ).
More precisely, it is now known since 1962 that there exist functions fk,l and a function g
P2n
in C(R) such that f (x1 , . . . , xn ) = k=0 g(fk,1 (x1 ) + · · · + fk,n xn ). As one can write finite
sums using functions of two variables like h(x, y) = x + y or h(x + y, z) = x + y + z two
variables suffice. The above form was given by by George Lorentz in 1962. Andrei Kolmogorov
reduced the problem in 1956 to functions of three variables. Vladimir Arnold showed then (as a
student at Moscow State university) in 1957 that one can do with two variables. The problem
came from a more specific problem in algebra, the problem of finding roots of a polynomial
p(x) = xn + a1 xn−1 + · · · an using radicals and arithmetic operations in the coefficients is not
possible in general for n ≥ 5. Erland Samuel Bring shows in 1786 that a quintic can be reduced
to x5 + ax + 1. In 1836 William Rowan Hamilton showed that the sextic can be reduced to
x6 + ax2 + bx + 1 to x7 + ax3 + bx2 + cx + 1 and the degree 8 to a 4 parameter problem
x8 + ax4 + bx3 + cx2 + dx + 1. Hilbert conjectured that one can not do better. They are the
Hilbert’s 13th problem, the sextic conjecture and octic conjecture. In 1957, Arnold and
Kolmogorov showed that no topological obstructions exist to reduce the number of variables.
Important progress was done in 1975 by Richard Brauer. Some history is given in [142]:
81. Determinants
The determinant of a n × n matrix A is defined as the sum π (−1)sign(π) A1π(1) · · · Anπ(n) ,
P
where the sum is over all n! permutations π of {1, . . . , n} and sign(π) is the signature of
the permutation π. The determinant functional satisfies the product formula det(AB) =
det(A)det(B). As the determinant is the constant coefficient of the characteristic polyno-
mial pA (x) = det(A − x1) = p0 (−x)n + p1 (−x)n−1 + · · · + pk (−x)n−k + · · · + pn of A, one can
get the coefficients of the product F T G of two n × m matrices F, G as follows:
P
Theorem: pk = |P |=k det(FP ) det(GP ).
The right hand side is a sum over all minors of length k including the empty one |P | =
T
P
0, where det(FP )Pdet(GP ) = 1. This implies det(1 + F G) = P det(FP ) det(GP ) and so
2
det(1 + F T F ) = PP det (F P ). The classical Cauchy-Binet theorem is the special case k = m,
T
where det(F G) = P det(FP )det(GP ) is a sum over all m×m patterns if n ≥ m. It has as even
more special case the Pythagorean consequence det(AT A) = P det(A2P ). The determinant
P
product formula is the even more special case when n = m. [215, 251, 201].
82. Triangles
A triangle T on a two-dimensional surface S is defined by three points A, B, C joined by three
geodesic paths. (It is assumed that the three geodesic paths have no self-intersections nor other
intersections besides A, B, C so that T is a topological disk with a piecewise geodesic boundary).
If α, β, γ are the inner anglesR of a triangle T located on a surface with curvature K, there
is the Gauss-Bonnet formula S K(x)dA(x) = χ(S), where dA denotes the area element on
34
OLIVER KNILL
the surface. This implies a relation between the integral of the curvature over the triangle and
the angles:
R
Theorem: α + β + γ = T K dA + π
This can be seen as a special

R Gauss-Bonnet result for Riemannian manifolds with bound-
ary as it is equivalent to T K dA + α + β + γ 0 = 2π with complementary angles α0 =
0
π − α, β 0 = π − β, γ 0 = π − γ. One can think of the vertex contributions as boundary cur-

vatures (generalized function). In the case of constant curvature K, the formula becomes
α + β + γ = KA + π, where A is the area of the triangle. Since antiquity, one knows the
flat case K = 0, where π = α + β + γ taught in elementary school. On the unit sphere
this is α + β + γ = A + π, result of Albert Girard which was predated by Thomas Harriot.
In the Poincaré disk model K = −1, this is α + β + γ = −A + π which is usually stated
that the area of a triangle in the disk is π − α − β − γ. This was proven by Johann Heinrich
Lambert. See [65] for spherical geometry and [15] for hyperbolic geometry, which are both part
of non-Euclidean geometry and now part of Riemannian geometry. [41, 223]
83. KAM
An area preserving map T (x, y) = (2x−y +cf (x), x) has an orbit (xn+1 , xn ) on T2 = (R/Z)2
which satisfies the recursion xn+1 − 2xn + xn−1
R 1 = cf (xn ). The 1-periodic function f is assumed
to be real-analytic, non-constant satisfying 0 f (x) dx = 0. In the case f (x) = sin(2πx), one
has the Standard map. When looking for invariant curves (q(t + α), q(t)) with smooth q, we
seek a solution of the nonlinear equation F (q) = q(t + α) − 2q(t) + q(t − α) − cf (q(t)) = 0. For
c = 0, there is the solution q(t) = t. The linearization dF (q)(u) = Lu = u(t + α) − 2u(t) +
u(t − α) − cf 0 (q(t))u(t) is a bounded linear operator on L2 (T) but not invertible for c = 0 so
that the implicit function theorem does not apply. The map Lu = u(t+α)−2u(t)+u(t−α)
becomes after a Fourier transform the diagonal matrix L̂ûn = [2 cos(nα) − 2]ûn which has the
inverse diagonal entries [2 cos(nα) − n]−1 leading to small divisors. A real number α is called
Diophantine if there exists a constant C such that for all integers p, q with q 6= 0, we have
|α−p/q| ≥ C/q 2 . KAM theory assures that the solution q(t) = t persists and remains smooth
if c is small. With solution the theorem means a smooth solution. For real analytic F , it
can be real analytic. The following result is a special case of the twist map theorem.
Theorem: For Diophantine α, there is a solution of F (q) = 0 for small |c|.
The KAM theorem was predated by the Poincaré-Siegel theorem in complex dynamics
which assured that if f is analytic near z = 0 and f 0 (0) = λ = exp(2πiα) with Diophantine
α, then there exists u(z) = z + q(z) such that f (u(z)) = u(λz) holds in a small disk 0: there
is an analytic solution q to the Schröder equation λz + g(z + q(z)) = q(λz). The question
about the existence of invariant curves is important as it determines the stability. The twist
map theorem result follows also from a strong implicit function theorem initiated by John
Nash and Jürgen Moser. For larger c, or non-Diophantine α, the solution q still exists but it is
no more continuous. This is Aubry-Mather theory. For c 6= 0, the operator L̂ is an almost
periodic Toeplitz matrix on l2 (Z) which is a special kind of discrete Schrödinger operator.
The decay rate of the off diagonals depends on the smoothness of f . Getting control of the
inverse can be technical [51]. Even in the Standard map case f (x) = sin(x), the composition
f (q(t)) is no more a trigonometric polynomial so that L̂ appearing here is not a Jacobi matrix
35
in a strip. The first breakthrough of the theorem in a frame work of Hamiltonian differential
equations was done in 1954 by Andrey Kolmogorov. Jürgen Moser proved the discrete twist
map version and Vladimir Arnold in 1963 proved the theorem for Hamiltonian systems. The
above stated result generalizes to higher dimensions where one looks for invariant tori called
KAM tori. one needs some non-degeneracy conditions See [74, 307, 308].
84. Continued Fraction

Given a positive square free integer d, the Diophantine equation√x2 −dy 2 = 1 is called √ Pell’s
equation.
√ Solving it means to find a nontrivial unit in the ring Z[ d] because (x + y d)(x −
y d) = 1. The trivial solutions are x = ±1, y = 0. Solving the equation is therefore part of the
Dirichlet unit problem from √ algebraic number theory. Let [a0 ; a1 , . . . ] denote the continued
fraction expansion of x = d. This means a0 = [x] is the integer part and [1/(x − a0 )] = a1
etc. If x = [a0 ; a1 , . . . , an + bn ], then an+1 = [1/bn ]. Let
√ pn /qn = [a0 ; a1 , a2 , . . . , an ] denote the
n’th convergent to the regular continued fraction of d. A solution (x1 , y1 ) which minimizes
x is called the fundamental solution. The theorem tells that it is of the form (pn , qn ):
Theorem: Any solution to the Pell’s equation is a convergent pn /qn .

√
One can find more solutions recursively because the ring of units in Z[ d] √ is Z2 × Cn for
√ some
cyclic group Cn . The other solutions (xk , yk ) can be obtained from xk + dyk = (x1 + dy1 )k .
One of the first instances, where the equation appeared is in the Archimedes cattle problem
which is x2 − 410286423278424y 2 = 1. The equation is named after John Pell, who has nothing
to do with the equation. It was Euler who attributed the solution by mistake to Pell. It was
first found by William Brouncker. The approach through continued fractions started with Euler
and Lagrange. See [343, 59, 278].
85. Gauss-Bonnet-Chern
Let (M, g) be a Riemannian manifold of dimension d with volume element dµ. If
ij
Rkl is Riemann curvature tensor with respect to the metric g, define the constant C =
σ(1)σ(2) σ(d−1)σ(d)
((4π)d/2 (−2)d/2 (d/2)!)−1 and the curvature K(x) = C σ,π sign(σ)sign(π)Rπ(1)π(2) · · · Rπ(d−1)π(d) ,
P
where the sum is over all permutations π, σ of {1, . . . , d}. It can be interpreted as a Pfaffian.
In odd dimensions, the curvature is zero. Denote by χ(M ) the Euler characteristic of M .
R
Theorem: M
K(x) dµ(x) = 2πχ(M ).
The case d = 2 was solved by Carl Friedrich Gauss and by Pierre Ossian Bonnet in 1848. Gauss
knew the theorem but never published it. In the case d = 2, the curvature K is the Gaussian
curvature which is the product of the principal curvatures κ1 , κ2 at a point. For a sphere
2
of radius R for example, the Gauss curvature is 1/RR and χ(M ) = 2. The volume form is
then the usual area element normalized so that M 1 dµ(x) = 1. Allendoerfer-Weil in 1943
gave the first proof, based on previous work of Allendoerfer, Fenchel and Weil. Chern finally,
in 1944 proved the theorem independent of an embedding. See [105], which features a proof of
Vijay Kumar Patodi.
36
OLIVER KNILL
86. Atiyah-Singer
Assume M is a compact orientable finite dimensional manifold of dimension n and assume D
is an elliptic differential operator D : E → F between two smooth vector bundles E, F
over M . Using multi-index notation Dk = ∂xk11 · · · ∂xknn , a differentialPoperator k ak (x)Dk x
P
is called elliptic if for all x, its symbol the polynomial σ(D)(y) = |k|=n ak (x)y k is not zero
for nonzero y. Elliptic regularity assures that both the kernel of D and the kernel of the
adjoint D∗ : F → E are both finite dimensional. The analytical index of D is defined as
χ(D) = dim(ker(D)) − dim(ker(D∗ )). We think of it as the Euler characteristic of D. The
topological index of D is defined as the integral of the n-form KD = (−1)n ch(σ(D))·td(T M ),
over M . This n-form is the cup product · of the Chern character ch(σ(D)) and the Todd
class of the complexified tangent bundle T M of M . We think about KD as a curvature.
Integration is done over the fundamental class [M ] of M which is the natural volume form
on M . The Chern character and the Todd classes are both mixed rational cohomology classes.
On a complex vector bundle E they are both given by concrete power series of Chern classes
ck (E) like ch(E) = ea1 (E) + · · · + ean (E) and td(E) = a1 (1 + e−a1 )−1 · · · an (1 + e−an )−1 with
ai = c1 (Li ) if E = L1 ⊕ · · · ⊕ Ln is a direct sum of line bundles.
R
Theorem: The analytic index and topological indices agree: χ(D) = M KD .
In the case when D = d+d∗ from the vector bundle of even forms E to the vector bundle of odd
forms F , then KD is the Gauss-Bonnet curvature and χ(D) = χ(M ). Israil Gelfand conjectured
around 1960 that the analytical index should have a topological description. The Atiyah-Singer
index theorem has been proven in 1963 by Michael Atiyah and Isadore Singer. The result
generalizes the Gauss-Bonnet-Chern and Riemann-Roch-Hirzebruch theorem. According to
[346], “the theorem is valuable, because it connects analysis and topology in a beautiful and
insightful way”. See [322].
87. Complex multiplication

A n’th root of unity is a solution to the equation z n = 1 in the complex plane C. It is called
primitive if it is not a solution to z k = 1 for some 1 ≤ k < n. A cyclotomic field is a
number field Q(ζn ) which is obtained by adjoining a complex primitive root of unity ζn to
Q. Every cyclotomic field is an Abelian field extension of the field of rational numbers Q. The
Kronecker-Weber theorem reverses this. It is also called the main theorem of class field
theory over Q
Theorem: Every Abelian extension L/Q is a subfield of a cyclotomic field.
Abelian field extensions of Q are also called class fields. It follows that any algebraic number
field K/Q with Abelian Galois group has a conductor, the smallest n such that K lies in
the field generated by n’th roots of unity. Extending this theorem to other base number fields is
Kronecker’s Jugendtraum or Hilbert’s twelfth problem. The theory of complex mul-
tiplication does the generalization for imaginary quadratic fields. The theorem was stated
by Leopold Kronecker in 1853 and proven by Heinrich Martin Weber in 1886. A generalization
to local fields was done by Jonathan Lubin and John Tate in 1965 and 1966. (A local field is
a locally compact topological field with respect to some non-discrete topology. The list of local
fields is R, C, field extensions of the p-adic numbers Qp , or formal Laurent series Fq ((t)) over
a finite field Fq .) The study of cyclotomic fields came from elementary geometric problems
37
like the construction of a regular n-gon with ruler and compass. Gauss constructed a regular
17-gon and showed that a regular n-gon can be constructed if and only if n is a Fermat
n
prime Fn = 22 + 1 (the known ones are 3, 6, 17, 257, 65537 and a problem of Eisenstein of 1844
asks whether there are infinitely many). Further interest came in the context of Fermat’s last
theorem because xn + y n = z n can be written as xn + y n = (x + y)(x + ζy) + · · · (x + ζ n−1 y),
where ζ is a n’th root of unity.
88. Choquet theory

Let K be a compact and convex set in a Banach space X. A point x ∈ K is called extreme
if x is not in an open interval (a, b) with a, b ∈ K. Let E be the set of extreme points in K. The
Krein-Milman theorem, proven in 1940 by Mark Krein and David Milman, assures R that K
is the convex hull of E. Given a probability measure µ on E, it defines the point x = ydµ(y).
We say that x is the Barycenter of µ. The Choquet theorem is
Theorem: Every point in K is a Barycenter of its extreme points.
This result of Choquet implies the Krein-Milman theorem. It generalizes to locally compact
topological spaces. The measure µ is not unique in general. It is in finite dimensions if K is
a simplex. But in general, as shown by Heinz Bauer in 1961, for an extreme point x ∈ K the
measure µx is unique. It has been proven by Gustave Choquet in 1956 and was generalized
by Erret Bishop and Karl de Leeuw in 1959. [328]
89. Helly’s theorem

Given a family K = {K1 , . . . Kn } of convex sets K1 , K2 , . . . , Kn in the Euclidean space Rd
and assume that n > d. Let Km denote the set of subsets of K which have exactly m elements.
We say that Km has the intersection property if every of its elements has a non-empty
common intersection. The theorem of Helly assures that
Theorem: Kn has the intersection property if Kd+1 has.
The theorem was proven in 1913 by Eduard Helly. It generalizes to an infinite collection
of compact, convex subsets. This theorem led Johann Radon to prove in 1921 the Radon
theorem which states that any set of d + 2 points in Rd can be partitioned into two disjoint
subsets whose convex hull intersect. A nice application of Radon’s theorem is the Borsuk-
Ulam theorem which states that a continuous function f from the d-dimensional sphere S n
to Rd must some pair of antipodal points to the same point: f (x) = f (−x) has a solution.
For example, if d = 2, this implies that on earth, there are at every moment two antipodal
points on the Earth’s surface for which the temperature and the pressure are the same. The
Borsuk-Ulam theorem appears first have been stated in work of Lazar Lyusternik and Lev
Shnirelman in 1930, and proven by Karol Borsuk in 1933 who attributed it to Stanislav Ulam.
90. Weak Mixing

An automorphism T of a probability space (X, A, m) is a measure preserving invertible
measurable transformation from X to X. It is called ergodic if T (A) = A implies m(A) = 0
or m(A) = 1. It is called mixing Pif m(T n (A) ∩ B) → m(A) · m(B) for n → ∞ for all A, B.
−1 n−1 k
It is called weakly mixing if n k=0 |m(T (A) ∩ B) − m(A) · m(B)| → 0 for all A, B ∈ A
38
OLIVER KNILL
and n → ∞. This is equivalent to the fact that the unitary operator U f = f (T ) on L2 (X) has
no point spectrum when restricted to the orthogonal complement of the constant functions.
A topological transformation (a continuous map on a locally compact topological space) with
a weakly mixing invariant measure is not integrable as for integrability, one wants every
invariant measure to lead to an operator U with pure point spectrum and conjugating it so to
a group translation. Let G be the complete topological group of automorphisms of (X, A, m)
with the weak topology: Tj converges to T weakly, if m(Tj (A)∆T (A)) → 0 for all A ∈ A; this
topology is metrizable and completeness is defined with respect to an equivalent metric.
Theorem: A generic T is weakly mixing and so ergodic.
Anatol Katok and Anatolii Mikhailovich Stepin in 1967 [231] proved that purely singular con-
tinuous spectrum of U is generic. A new proof was given by [84] and a short proof in using
Rokhlin’s lemma, Halmos conjugacy lemma and a Simon’s “wonderland theorem” estab-
lishes both genericity of weak mixing and genericity of singular spectrum. On the topological
side, a generic volume preserving homeomorphism of a manifold has purely singular continuous
spectrum which strengthens Oxtoby-Ulam’s theorem [321] about generic ergodicity. [232, 181]
The Wonderland theorem of Simon [369] also allowed to prove that a generic invariant measure
of a shift is singular continuous [245] or that zero-dimensional singular continuous spectrum is
generic for open sets of flows on the torus allowing also to show that open sets of Hamiltonian
systems contain generic subset with both quasi-periodic as well as weakly mixing invariant tori
[246]
91. Universality
The space X of unimodular maps is the set of twice continuously differentiable even maps
f : [−1, 1] → [−1, 1] satisfying f (0) = 1 f 00 (x) < 0 and λ = f (1) < 0. The Feigenbaum-
Cvitanović functional equation (FCE) is g = T g with T (g)(x) = λ1 g(g(λx)). The map T is
a renormalization map.
Theorem: There exists an analytic hyperbolic fixed point of T .
The first proof was given by Oscar Lanford III in 1982 (computer assisted). See [212, 213].
That proof also established that the fixed point is hyperbolic with a one-dimensional unstable
manifold and positive expanding eigenvalue. This explains some universal features of uni-
modular maps found experimentally in 1978 by Mitchell Feigenbaum and which is now called
Feigenbaum universality. The result has been ported to area preserving maps [124].
92. Compactness
Let X be a compact metric space (X, d). The Banach space C(X) of real-valued continuous
functions is equipped with the supremum norm. A closed subset F ⊂ C(X) is called uniformly
bounded if for every x the supremum of all values f (x) with f ∈ F is bounded. The set F is
called equicontinuous if for every x and every > 0 there exists δ > 0 such that if d(x, y) < δ,
then |f (x) − f (y)| < for all f ∈ F . A set F is called precompact if its closure is compact.
The Arzelà-Ascoli theorem is:
Theorem: Equicontinuous uniformly bounded sets in C(X) are precompact.
39
The result also holds on Hausdorff spaces and not only metric spaces. In the complex, there
is a variant called Montel’s theorem which is the fundamental normality test for holomorphic
functions: an uniformly bounded family of holomorphic functions on a complex domain G is
normal meaning that its closure is compact with respect to the compact-open topology.
The compact-open topology in C(X, Y ) is the topology defined by the sub-base of all contin-
uous maps fK,U : f : K → U , where K runs over all compact subsets of X and U runs over all
open subsets of Y .
93. Geodesic
The geodesic distance d(x, y) between two points x, y on a Riemannian manifold (M, g) is
defined as the length of the shortest geodesic γ connecting x with y. This renders the manifold
a metric space (M, d). We assume it is locally compact, meaning that every point x ∈ M has
a compact neighborhood. A metric space is called complete if every Cauchy sequence in M
has a convergent subsequence. (A sequence xk is called a Cauchy sequence if for every > 0,
there exists n such that for all i, j > n one has d(xi , xj ) < .) The local existence of differential
equations assures that the geodesic equations exist for small enough time. This can be restated
that the exponential map v ∈ Tx M → M assigning to a point v 6= 0 in the tangent space
Tx M the solution γ(t) with initial velocity v/|v| and t ≤ |v|, and γ(0) = x. A Riemannian
manifold M is called geodesically complete if the exponential map can be extended to the
entire tangent space Tx M for every x ∈ M . This means that geodesics can be continued for all
times. The Hopf-Rinov theorem assures:
Theorem: Completeness and geodesic completeness are equivalent.
The theorem was named after Heinz Hopf and his student Willi Rinov who published it in 1931.
See [117].
94. Crystallography
A wall paper group is a discrete subgroup of the Euclidean symmetry group E2 of the
plane. Wall paper groups classify two-dimensional patterns according to their symmetry. In
the plane R2 , the underlying group is the group E2 of Euclidean plane symmetries which
contain translations rotations or reflections or glide reflections. This group is the group
of rigid motions. It is a three dimensional Lie group which according to Klein’s Erlangen
program characterizes Euclidean geometry. Every element in E2 can be given as a pair
(A, b), where A is an orthogonal matrix and b is a vector. A subgroup G of E2 is called discrete
if there is a positive minimal distance between two elements of the group. This implies the
crystallographic restriction theorem assuring that only rotations of order 2, 3, 4 or 6 can
appear. This means only rotations by 180, 120, 90 or 60 degrees can occur in a Wall paper
group.
Theorem: There are 17 wallpaper groups
The first proof was given by Evgraf Fedorov in 1891 and then by George Polya in 1924. in
three dimensions there are 230 space groups and 219 types if chiral copies are identified. In
space there are 65 space groups which preserve the orientation. See [320, 177, 221].
40
OLIVER KNILL
95. Quadratic forms
Pn matrix Q of size n × n with integer entries defines a integer quadratic

A symmetric square
form Q(x) = i,j=1 Qij xi xj . It is called positive if Q(x) > 0 whenever x 6= 0. A positive
integral quadratic form is called universal if its range is N. For example, by the Lagrange
four square theorem, the form Q(x1 , x2 , x3 , x4 ) = x21 + x22 + x23 + x24 is universal. The
Conway-Schneeberger fifteen theorem tells
Theorem: Q is universal if it has {1, . . . 15} in the range.
The interest in quadratic forms started in the 17’th century especially about numbers which
can be represented as sums x2 + y 2 . Lagrange, in 1770 proved the four square theorem. In
1916, Ramajujan listed all diagonal quaternary forms which are universal. The 15 theorem was
proven in 1993 by John Conway and William Schneeberger (a student of Conway’s in a graduate
course given in 1993). There is an analogue theorem for integral positive quadratic forms,
these are defined by positive definite matrices Q which take only integer values. The binary
quadratic form x2 + xy + y 2 for example is integral but not an integer quadratic form because
the corresponding matrix Q has fractions 1/2. In 2005, Bhargava and Jonathan Hanke proved
the 290 theorem, assuring that an integral positive quadratic form is universal if it contains
{1, . . . , 290} in its range. [92].
96. Sphere packing

A sphere packing in Rd is an arrangement of non-overlapping unit spheres in the d-dimensional
Euclidean space Rd with volume measure µ. It is known since [173] that packings with maximal
d
S the ball of radius r centered at x ∈ R . If X is the set of
densities exist. Denote by Br (x)
centers of the sphere and P = x∈X B1 (x) is the union of the unit
R balls centered
R at points in
X, then the density of the packing is defined as ∆d = lim sup Br (0) P dµ/ Br (0) 1 dµ. The
sphere packing problem is now solved in 5 different cases:
Theorem: Optimal sphere packings are known for d = 1, 2, 3, 8, 24.
√
The one-dimensional case ∆1 = 1 is trivial. The case ∆2 = π/ 12 was known since Axel Thue
in 1910 but proven only by Lásló Fejes Toóth in 1943.√The case d = 3 was called the Kepler
conjecture as Johannes Kepler conjectured ∆3 = π/ 18. It was settled by Thomas Hales in
1998 using computer assistance. A complete formal proof appeared in 2015. The case d = 8 was
settled by Maryna Viazovska who proved in 2017 [411] that ∆8 = π 4 /384 and also established
uniqueness. The densest packing in the case d = 8 is the E8 lattice. The proof is based on
linear programming bounds developed by Henry Cohn and Noam Elkies in 2003. Later with
other collaborators, she also covered the case d = 24. the densest packing in dimension 24 is
the Leech lattice. For sphere packing see [99, 98].
97. Sturm theorem

Given a square free real-valued polynomial p let pk denote the Sturm chain, p0 = p, p1 = p0 ,
p2 = p0 mod p1 , p3 = p1 mod p2 etc. Let σ(x) be the number of sign changes ignoring zeros
in the sequence p0 (x), p1 (x), . . . , pm (x).
Theorem: The number of distinct roots of p in (a, b] is σ(b) − σ(a).
41
Sturm proved the theorem in 1829. He found his theorem on sequences while studying solutions
of differential equations Sturm-Liouville theory and credits Fourier for inspiration. See [337].
98. Smith Normal form

A integer m × n matrix A is said to be expressible in Smith normal form if there exists an
invertible m × m matrix S and an invertible n × n matrix T so that SM T is a diagonal matrix
Diag(α1 , . . . , αr , 0, 0, 0) with αi |αi+1 . The integers αi are called elementary divisors. They
can be written as αi = di (A)/di−1 (A), where d0 (A) = 1 and dk (A) is the greatest common
divisor of all k × k minors of A. The Smith normal form is called unique if the elementary
divisors αi are determined up to a sign.
Theorem: Any integer matrix has a unique Smith normal form.
The result was proven by Henry John Stephen Smith in 1861. The result holds more generally
in a principal ideal domain, which is an integral domain (a ring R in which ab = 0 implies
a = 0 or b = 0) in which every ideal (an additive subgroup I of the ring such that ab ∈ I if
a ∈ I and b ∈ R) is generated by a single element.
99. Spectral perturbation

A complex valued matrix A is self-adjoint = Hermitian if A∗ = A, where A∗ij = Aji . The
spectral theorem assures that A has real eigenvalues Given two selfadjoint complex n × n
matrices A, B with eigenvalues α1 ≤ α2 ≤ · · · ≤ αn and β1 ≤ β2 ≤ · · · ≤ βn , one has the
Lidskii-Last theorem:
Pn Pn
Theorem: j=1 |αj − βj | ≤ i,j=1 |A − B|ij .
The result has been deduced

P by Yoram P Last (around 1993) from Lidskii’s inequality found
in 1950 by Victor Lidskii j |αj − βj | ≤ j |γj | where γj are the eigenvalues of C = B − A (see
p p
P P
[370] page 14). The original Lidskii inequality also holds for p ≥ 1: j |αj − βj | ≤ j |γj | .
Last’s spin on it allows to estimate the l1 spectral distance of two self-adjoint matrices using
the l1 distance of the matrices. This is handy as we often know the matrices A, B explicitly
rather than the eigenvalues γj of A − B.
100. Radon transform

In order to solve the tomography problem like magnetic resonance imaging (MRI) of
finding the density function g(x, y, z) of a three dimensional body, one looks at a slice f (x, y) =
g(x,
R y, c), where z = c is kept constant and measures the Radon transform R(f )(p, θ) =
{x cos(θ)+y sin(θ)=p}
f (x, y) ds. This quantity is the absorption rate due to nuclear magnetic
resonance along the line L of polar angle α in distance p from the center. Reconstructing
f (x, y) = g(x, y, c) for different c allows to recover the tissue density g and so to “see inside
the body”.
Theorem: The Radon transform can be diagonlized and so pseudo inverted.
42
OLIVER KNILL
We only need that the Fourier series f (r, φ) = n fn (r)einφPconverges

P
P uniformly for all r > 0 and
that fn (r) has a Taylor series. The expansion f (r, φ) = n∈Z ∞ k=1 fn,k ψn,k with ψn,k (r, φ) =
π/2
r−k einφ is an eigenfunction expansion with eigenvalues λn,k = 2 0 cos(nx) cos(x)(k−1) dx =
R
π Γ(k+1)
2k−1 ·k
· Γ( k+n+1 )Γ( k−n+1 )
. The inverse problem is subtle due to the existence of a kernel
2 2
spanned by {ψn,k | (n + k) odd , |n| > k}. One calls it an ill posed problem in the sense of
Hadamard. The Radon transform was first studied by Johann Radon in 1917 [194].
101. Linear programming

Given two vectors c ∈ Rm and b ∈ Rn , and a n × m matrix A, a linear program is the
variational problem on Rm to maximize f (x) = c · x subject to the linear constraints Ax ≤ b
and x ≥ 0. The dual problem is to minimize b · y subject to to AT y ≥ c, y ≥ 0. The maximum
principle for linear programming is tells that the solution is on the boundary of the convex
polytop formed by the feasable region defined by the constraints.
Theorem: Local optima of linear programs are global and on the boundary
.
Since the solutions are located on the vertices of the polytope defined by the constraints the
simplex algorithm for solving linear programs works: start at a vertex of the polytop, then
move along the edges along the gradient until the optimum is reached. If A = [2, 3] and
x = [x1 , x2 ] and b = 6 and c = [3, 5] we have n = 1, m = 2. The problem is to maximize
f (x1 , x2 ) = 3x1 + 5x2 on the triangular region 2x1 + 3x2 ≤ 6, x1 ≥ 0, x2 ≥ 0. Start at (0, 0),
the best improvement is to go to (0, 2) which is already the maximum. Linear programming is
used to solve practical problems in operations research. The simplex algorithm was formulated
by George Dantzig in 1947. It solves random problems nicely but there are expensive cases in
general and it is possible that cycles occur. One of the open problems of Steven Smale asks
for a strongly polynomial time algorithm deciding whether a solution of a linear programming
problem exists. [310]
102. Random Matrices

A random matrix A is given by an n × n array of independent, identically distributed
√ random
variables Aij of zero mean
P and standard deviation 1. The eigenvalues λj of A/ n define a
discrete measure µn = j δλj called spectral measure of A. The circular law on the
complex plane C is the probability measure µ0 = 1D /π, where D = {|z| ≤ 1} is the unit disk.
A sequence νn of probability measures converges
R weakly orR in law to ν if for every continuous
and bounded function f : C → C one has f (z) dνn (z) → f (z) dν(z). The circular law is:
Theorem: Almost surely, the spectral measures converge µn → µ0 .
One can think of An as a sequence of larger and larger matrix valued random variables. The
circular law tells that the eigenvalues fill out the unit disk in the complex plane uniformly when
taking larger and larger matrices. It is a kind of central limit theorem. An older version due
to Eugene Wigner from 1955 is the semi-circular law telling that √ in the self-adjoint case, the
now real measures µn converge to a distribution with density 4 − x2 /(2π) on [−2, 2]. The
circular law was stated first by Jean Ginibre in 1965 and Vyacheslav Girko 1984. It was proven
first by Z.D. Bai in 1997. Various authors have generalized it and removed more and more
43
moment conditions. The latest condition was removed by Terence Tao and Van Vu in 2010,
proving so the above “fundamental theorem of random matrix theory”. See [401].
103. Diffeomorphisms
Let M be a compact Riemannian surface and T : M → M a C 2 -diffeomorphism. A Borel
probability measure µ on M is T -invariant if µ(T (A)) = µ(A) for all A ∈ A. It is called
ergodic if T (A) = A implies µ(A) = 1 or µ(A) = 0. The Hausdorff dimension dim(µ) of
a measure µ is defined as the Hausdorff dimension of the smallest Borel set A of full measure
µ(A) = 1. The entropy hµ (T ) is the Kolmogorov-Sinai entropy of the measure-preserving
dynamical system (X, T, µ). For an ergodic surface diffeomorphism, the Lyapunov exponents
λ1 , λ2 of (X, T, µ) are the logarithms of the eigenvalues of A = limn→∞ [(dT n (x))∗ dT n (x)]1/(2n) ,
which is a limiting Oseledec matrix and constant µ almost everywhere due to ergodicity. Let
λ(T, µ) denote the Harmonic mean of λ1 , −λ2 . The entropy-dimension-Lyapunov theorem
tells that for every T -invariant ergodic probability measure µ of T , one has:
Theorem: hµ = dim(µ)λ/2.
This formula has become famous because it relates “entropy”, “fractals” and “chaos”, which are
all “rock star” notions also outside of mathematics. The theorem implies in the case of Lebesgue
measure preserving symplectic transformation, where dim(µ) = 2 and λ1 = −λ2 that “entropy
= Lyaponov exponent” which is a formula of Pesin given by hµ (T ) = λ(T, µ). A similar result
holds for circle diffeomorphims or smooth interval maps, where hµ (T ) = dim(µ)λ(T, µ). The
notion of Hausdorff dimension was introduced by Felix Hausdoff in 1918. Entropy was defined
in 1958 by Nicolai Kolmogorov and in general by Yakov Sinai in 1959, Lyapunov exponents
were introduced with the work of Valery Oseledec in 1965. The above theorem is due to Lai-
Sang Young who proved it in 1982. P Francois Ledrapier and Lai-Sang Young proved in 1985
that in arbitrary dimensions, hµ = j λj γj , where γj are dimensions of µ in the direction
of the Oseledec spaces Ej . This is called P the+ Ledrappier-Young formula. It implies the
+
Margulis-Ruelle inequality hµ (T ) ≤ j λj (T ), where λj = max(λj , 0) and λj (T ) are the
Lyapunov exponents. In the case of a smooth P T+-invariant measure µ or more generally, for
SRB measures, there is an equality hµ (T ) = j λj (T ) which is called the Pesin formula. See
[230, 125].
104. Linearization
If F : M → M is a globally Lipschitz continuous function on a finite dimensional vector space
M , then the differential equation x0 = F (x) has a global solution x(t) = f t (x(0)) (a local
by Picard-Lindelöf ’s existence theorem and global by the Grönwall inequality). An
equilibrium point of the system is a point x0 for which F (x0 ) = 0. This means that x0 is a
fixed point of a differentiable mapping f = f 1 , the time-1-map. We say that f is linearizable
near x0 if there exists a homeomorphism φ from a neighborhood U of x0 to a neighborhood V of
x0 such that φ ◦ f ◦ φ−1 = df . The Sternberg-Grobman-Hartman linearization theorem
is
Theorem: If f is hyperbolic, then f is linearizable near x0 .
The theorem was proven by D.M. Grobman in 1959 Philip Hartman in 1960 and by Shlomo
Sternberg in 1958. This implies the existence of stable and unstable manifolds passing
44
OLIVER KNILL
through x0 . One can show more and this is due to Sternberg who wrote a series of papers
starting 1957 [385]: if A = df (x0 ) satisfies no resonance condition meaning that no relation
λ0 = λ1 · · · λj exists between eigenvalues of A, then a linearization to order n is a C n map
φ(x) = x + g(x), with g(0) = g 0 (0) = 0 such that φ ◦ f ◦ φ−1 (x) = Ax + o(|x|n ) near x0 . We
say then that f can be n-linearized near x0 . The generalized result tells that non-resonance
fixed points of C n maps are n-linearizable near a fixed point. See [273].
105. Fractals
n
An iterated function system is a finite set of contractions {fi }P
i=1 on a complete metric space
(X, d). The corresponding Huntchingson operator H(A) = i fi (A) is then a contraction
on the Hausdorff metric of sets and has a unique fixed point called the attractor S of the
iteratedPfunction system. The definition of Hausdorff dimension is as follows: define hsδ (A) =
inf U ∈U i |Ui |s , where U is a δ-cover of A. And hs (A) = limδ→0 Hδs (A). The Hausdorff
dimension dimH (S) finally is the value s, where hs (S) jumps from ∞ to 0. If the contractions
are maps with contraction factors 0 < λj < 1 then the Hausdorff dimension of the attractor S
can be estimated with the the similarity dimension of Pnthe contraction vector (λ1 , . . . , λn ):
−s
this number is defined as the solution s of the equation i=1 λi = 1.
Theorem: dimhausdorff (S) ≤ dimsimilarity (S).
There is an equality if fi are all affine contractions like fi (x) = Ai λx + βi with the same
contraction factor and Ai are orthogonal and βi are vectors (a situation which generates a
large class of popular fractals). For equality one also has to assume that there is an open
non-empty set G such that Gi = fi (G) are disjoint. In the case λj = λ are all the same
then nλ−dim = 1 which implies dim(S) = − log(n)/ log(λ). For the Smith-Cantor set S,
where f1 (x) = x/3 + 2/3, f2 (x) = x/3 and G = (0, 1). One gets with n = 2 and λ = 1/3
the dimension dim(S) = log(2)/ log(3). For the Menger carpet with n = 8 affine maps
fij (x, y) = (x/3 + i/3, y/3 + j/3) with 0 ≤ i ≤ 2, 0 ≤ j ≤ 2, (i, j) 6= (1, 1), the dimension is
log(8)/ log(3). The Menger sponge is the analogue object with n = 20 affine contractions
in R3 and has dimension log(20)/ log(3). For the Koch curve on the interval, where n = 4
affine contractions of contraction factor 1/3 exist, the dimension is log(4)/ log(3). These are
all fractals, sets with Hausdorff dimension different from an integer. The modern formulation
of iterated function systems is due to John E. Hutchingson from 1981. Michael Barnsley used
the concept for a fractal compression algorithms, which uses the idea that storing the rules
for an iterated function system is much cheaper than the actual attractor. Iterated function
systems appear in complex dynamics in the case when the Julia set is completely disconnected,
they have appeared earlier also in work of Georges de Rham 1957. See [284, 140].
106. Strong law of small numbers

Like the Bayes theorem or the Pigeon hole principle which both are too simple to qualify as
“theorems” but still are of utmost importance, the “Strong law of small numbers” is not really
a theorem but a fundamental mathematical principle. It is more fundamental than a
specific theorem as it applies throughout mathematics. It is for example important in Ramsey
theory: The statement is put in different ways like ”There aren’t enough small numbers to
meet the many demands made of them”. [178] puts it in the following catchy way:
45
Theorem: You can’t tell by looking.
The point was made by Richard Guy in [178] who states two “corollaries”: “superficial sim-
ilarities spawn spurious statements” and “early exceptions eclipse eventual essen-
tials”. The statement is backed up with countless many examples (a list of 35 are given in
n
[178]). Famous are Fermat’s claim that all Fermat primes 22 + 1 are prime or the claim that
the number π3 (n) of primes of the form 4k + 3 in {1, . . . , n} is larger than π1 (n) of primes of
the form 4k + 1 so that the 4k + 3 primes win the prime race. Hardy and Littlewood showed
however π3 (n) − π1 (n) changes sign infinitely often. The prime number theorem extended to
arithmetic progressions shows π1 (n) ∼ n/(2 log(n)) and π3 (n) ∼ n/(2 log(n)) but the density
of numbers with π3 (n) > π1 (n) is larger than 1/2. This is the Chebyshev bias. Experiments
then suggested the density to be 1 but also this is false: the density of numbers for which
π3 (n) > π1 (n) is smaller than 1. The principle is important in a branch of combinatorics called
Ramsey theory. But it not only applies in discrete mathematics. There are many examples,
where one can not tell by looking. When looking at the boundary of the Mandelbrot set for
example, one would tell that it is a fractal with Hausdorff dimension between 1 and 2. In reality
the Hausdorff dimension is 2 by a result of Mitsuhiro Shishikura. Mandelbrot himself thought
first “by looking” that the Mandelbrot set M is disconnected. Douady and Hubbard proved
M to be connected.
107. Ramsey Theory

Let G be the complete graph with n vertices. An edge labeling with r colors is an assignment
of r numbers to the edges of G. A complete sub-graph of G is called a clique. If it is has
s vertices, it is denoted by Ks . A graph G is called monochromatic if all edges in G have
the same color. (We use in here coloring as a short for edge labeling and not in the sense
of chromatology where an edge coloring assumes that intersecting edges have different colors.)
Ramsey’s theorem is:
Theorem: For large n, every r-colored Kn contains a monochromatic Ks .
So, there exist Ramsey numbers R(r, s) such that for n ≥ R(r, s), the edge coloring of one
of the s-cliques can occur. A famous case is the identity R(3, 3) = 6. Take n = 6 people. It
defines the complete graph G. If two of them are friends, color the edge blue, otherwise red.
This friendship graph therefore is a r = 2 coloring of G. There are 78 possible colorings. In
each of them, there is a triangle of friends or a triangle of strangers. In a group of 6 people,
there are either a clique with 3 friends or a clique of 3 complete strangers. The Theorem was
proven by Frank Ramsey in 1930. Paul Erdoes asked to give explicit estimated R(s) which is
the least integer n such that any graph on n vertices contains either a clique of size s (a set
where all are connected to each other) or an independent set of size s (a set where none are
connected to each other). Graham for example asks whether the limit R(n)1/n exists. Ramsey
theory also deals other sets: van der Waerden’s theorem from 1927 for example tells that
if the positive integers N are colored with r colors, then for every k, there exists an N called
W (r, k) such that the finite set {1 . . . , N } has an arithmetic progression with the same color.
For example, W (2, 3) = 9. Also here, it is an open problem to find a formula for W (r, k) or
even give good upper bounds. [169] [168]
46
OLIVER KNILL
108. Poincaré Duality

For a differentiable Riemannian n-manifold (M, g) there is an exterior derivative d = dp
which maps p-forms Λp to (p + 1)-forms Λp+1 . For p = 0, the derivative is called the gradient,
for p = 1, the derivative is called the curl and for p = d − 1, the derivative is the adjoint of
divergence. The Riemannian metric defines an inner product hf, hi on Λp allowing so to see
Λp as part of a Hilbert space and to define the adjoint d∗ of d. It is a linear map from Λp+1
to Λp . The exterior derivative defines so the self-adjoint Dirac operator D = d + d∗ and the
Hodge Laplacian L = D2 = dd∗ + d∗ d which now leaves each Λp invariant. Hodge theory
assures that dim(ker(L|Λp )) = bp = dim(H p (M )), where H p (M ) are the p’th cohomology
group, the kernel of dp modulo the image of dp−1 . Poincaré duality is:
Theorem: If M is orientable n-manifold, then bk (M ) = bn−k (M ).
The Hodge dual of f ∈ Λp is defined as the unique ∗g ∈ Λn−p satisfying hf, ∗gi = hf ∧ g, ωi
where ω is the volume form. One has d∗ f = (−1)d+dp+1 ∗ d ∗ f and L ∗ f = ∗Lf . This implies
that ∗ is a unitary map from ker(L|Λp ) to ker(L|Λd−p ) proving so the duality theorem. For
n = 4k, one has ∗2 = 1, allowing to define the Hirzebruch signature σ := dim{u|Lu =
0, ∗u = u} − dim(u|Lu = 0, ∗u = −u}. The Poinaré duality theorem was first stated by Henri
Poincaré in 1895. It took until the 1930ies to clean out the notions and make it precise. The
Hodge approach establishing an explicit isomorphism between harmonic p and n − p forms
appears for example in [105].
109. Rokhlin-Kakutani approximation

Let T be an automorphism of a probability space (Ω, A, µ). This means µ(A) = µ(T (A)) for
all A ∈ A. The system T is called aperiodic, if the set of periodic points P = {x ∈
Ω | ∃n > 0, T n x = x} has measure µ(P ) = 0. A set B ∈ A which has the property that
B, T (B), . . . , T n−1 (B) are disjoint is called a Rokhlin tower. The measure of the tower is
µ(B ∪ · · · ∪ T n−1 (B)) = nµ(B). We call it an (1 − )-Rokhlin tower. We say T can be
approximated arbitrary well by Rokhlin towers, if for all > 0, there is an (1 − ) Rokhlin
tower.
Theorem: An aperiodic T can be approximated well by Rokhlin towers.
The result was proven by Vladimir Abramovich Rokhlin in his thesis 1947 and independently
by Shizuo Kakutani in 1943. The lemma can be used to build Kakutani skyscrapers, which
are nice partitions associated to a transformation. This lemma allows to approximate an
aperiodic transformation T by a periodic transformations Tn . Just change T on T n−1 (B) so
that Tnn (x) = x for all x. The theorem has been generalized by Donald Ornstein and Benjamin
Weiss to higher dimensions like Zd actions of measure preserving transformations where the
periodicity assumption is replaced by the assumption that the action is free: for any n 6= 0,
the set T n (x) = x has zero measure. See [100, 152, 181].
110. Lax approximation

On the group X of all measurable, invertible transformations on the d-dimensional torus
X = Td which preserve the Lebesgue volume measure, one has the metric
δ(T, S) = |δ(T (x), S(x))|∞ ,
47
where δ is the geodesic distance on the flat torus and where | · |∞ is the L∞ supremum norm.
Lets call (Td , T, µ) a toral dynamical system if T is a homeomorphism, a continuous
transformation with continuous inverse. A cube exchange transformation on Td is a peri-
odic,
Qd piecewise affine measure-preserving transformation T which permutes rigidly all the cubes
d
i=1 [ki /n, (ki + 1)/n], where ki ∈ {0, . . . , n − 1}. Every point in T is T periodic. A cube ex-
change transformation is determined by a permutation of the set {1, . . . , n}d . If it is cyclic,
the exchange transformation is called cyclic. A theorem of Lax [275] states that every toral
dynamical system can approximated in the metric δ by cube exchange transformations. The
approximations can even be cyclic [13].
Theorem: Toral systems can be approximated by cyclic cube exchanges
The result is due to Peter Lax [275]. The proof of this result uses Hall’s marriage theorem
in graph theory (for a ’book proof’ of the later theorem, see [9]). Periodic approximations of
symplectic maps work surprisingly well for relatively small n (see [339]). On the Pesin region
this can be explained in part by the shadowing property [230]. The approximation by cyclic
transformations make long time stability questions look different [180].
111. Sobolev embedding

R
All functions are defined on Rn , integrated over Rn and assumed R to be locally integrable
meaning that for every compact set K the Lebesgue integral K |f | dx is finite. For functions
in Cc∞ which serve as test functions, partial derivatives ∂i = ∂/∂xi and more general
differential operators
R Dk = ∂Rxk11 · · · ∂xknn can be applied. A function g is a weak partial
derivative
R of f if f ∂i φdx = − gφdx for all test functions φ. For p ∈ [1, ∞), the Lp space is
{f | |f |p dx < ∞}. The Sobolev space W k,p is the set of functions for which all k’th weak
derivatives are in Lp . So W 0,p = Lp . The Hölder space C r,α with r ∈ N, α ∈ (0, 1] is defined
as the set of functions for which all r’th derivatives are α-Hölder continuous. It is a Banach
space with norm max|k|≤r ||Dk f ||∞ + max|k|=r ||Dk f ||α , where ||f ||∞ is the supremum norm
and ||f ||α is the Hölder coefficient supx6=y |f (x) − f (y)|/|x − y|α . The Sobolev embedding
theorem is
Theorem: If n < p and l = r + α < k − n/p, one has W k,p ⊂ C r,α .
([371] states this as Theorem 6.3.6) gives some history: generalized functions appeared
first in the work of Oliver Heaviside in the form of “operational calculus. Paul Dirac used
the formalism in quantum mechanics. In the 1930s, Kurt Otto Friedrichs, Salomon Bocher
and Sergei Sobolev define weak solutions of PDE’s. Schwartz used the Cc∞ functions, smooth
functions of compact support. This means that the existence of k weak derivatives implies the
existence of actual derivatives. For p = 2, the spaces W k are Hilbert spaces and the theory a
bit simpler due to the availability of Fourier theory, where tempered distributions flourished.
In that case, one can define for any real s > 0 the Hilbert space H s as the subset of all f ∈ S 0
for which (1 + |ξ|2 )s/2 fˆ(ξ) is in L2 . The Schwartz test functions S consists of all C ∞ functions
having bounded semi norms ||φ||k = max|α|+|β|≤k ||xβ Dα φ||∞ < ∞ where α, β ∈ Nn . Since S is
larger than the set of smooth functions of compact support, the dual space S 0 is smaller. They
are tempered distributions. Sobolev emedding theorems like above allow to show that weak
solutions of PDE’s are smooth: for example, if the Poisson problem ∆f = V f with smooth V
is solved by a distribution f , then f is smooth. [62, 371]
48
OLIVER KNILL
112. Whitney embedding

A smooth n-manifold M is a metric space equipped with a cover Uj = φ−1 j (B) with B = {x ∈
n 2 −1 n 2
R | |x| < 1}) or Uj = φj (H) with H = {x ∈ R | |x| < 1, x0 ≥ 0}) with δH = {x ∈
H | x0 = 0} such that the homeomorphisms φj : Uj → B or φj : Uj → H lead to smooth
transition maps φkj = φj φ−1 k from φk (Uj ∩ Uk ) to φj (Uj ∩ Uk ) which have the property that all
restrictions of φkj from δφk (Uj ∩ Uk ) to δφj (Uj ∩ Uk ) are smooth too. The boundary δM of
M now naturally is a smooth (n − 1) manifold, the atlas being given by the sets Vj = φj (δH)
for the indices j which map φj : Uj → H. Two manifolds M, N are diffeomorphic if there
is a refinement {Uj , φj } of the atlas in M and a refinement {Vj , ψj } of the atlas in N such
that φj (Uj ) = ψj (Vj ). A manifold M can be smoothly embedded in Rk if there is a smooth
injective map f from M to Rk such that the image f (M ) is diffeomorphic to M .
Theorem: Any n-manifold M can be smoothly embedded in R2n .

The theorem has been proven by Hassler Whitney in 1926 who also was the first to give a
precise definition of manifold in 1936. The standard assumption is that M is second countable
Hausdorff but as every smooth finite dimensional manifold can be upgraded to be Riemannian,
the simpler metric assumption is no restriction of generality. The modern point of view is to
see M as a scheme over Euclidean n-space, more precisely as a ringed space, that is locally
the spectrum of the commutative ring C ∞ (B) or C ∞ (H). The set of manifolds is a category
in which the smooth maps M → N are the morphisms. The cover Uj defines an atlas and
the transition maps φj allow to port notions like smoothness from Euclidean space to M . The
maps φ−1 −1
j : B → M or φj : H → M parametrize the sets Uj . [426].
113. Artificial intelligence

Like meta mathematics or reverse mathematics, the field of artificial intelligence (AI)
can be considered as part of mathematics. It is related of data science (algorithms for data
mining, and statistics) computation theory (like complexity theory) language theory and
especially grammar and evolutionary dynamics, optimization problems (like solving op-
timal transport or extremal problems) solving inverse problems (like developing algorithms
for computer vision or optical character or speech recognition), cognitive science as well as
pedagogy in education (human or machine learning and human motivation). There is no
apparent “fundamental theorem” of AI, (except maybe Marvin Minsky’s ”The most efficient
way to solve a problem is to already know how to solve it.” [302], which is a surprisingly deep
statement as modern AI agents like Alexa, Siri, Google Home, IBM Watson or Cortana
demonstrate; they compute little, they just know or look up - or annoy you to look it up your-
self...). But there is a theorem of Lebowski on machine super intelligence which taps
into the uncharted territory of machine motivation
Theorem: No AI will bother after hacking its own reward function.
The picture [261] is that once the AI has figured out the philosophy of the “Dude” in the Cohen
brothers movie Lebowski, also repeated mischiefs does not bother it and it “goes bowling”.
Objections are brushed away with “Well, this is your, like, opinion, man”. Two examples
of human super intelligent units who have succeeded to hack their own reward function are
Alexander Grothendieck or Grigori Perelman. The Lebowski theorem is due to Joscha Bach
[26], who stated this theorem of super intelligence in a tongue-in-cheek tweet. From a
49
mathematical point of view, the smartest way to “solve” an optimal transport problem is to
change the utility function. On a more serious level, the smartest way to “solve” the continuum
hypothesis is to change the axiom system. This is a cheat, but on a meta level, more creativity
is possible. A precursor is Stanislav Lem’s notion of a mimicretin [277], a computer that plays
stupid in order, once and for all, to be left in peace or the machine in [6] who develops humor
and enjoys fooling humans with the answer to the ultimate question: ”42”. This document
entry is the analogue to the ultimate question: “What is the fundamental theorem of AI”?
114. Stokes Theorem

On a smooth orientable n-dimensional manifold M , one has Λp , the vector bundle of smooth
differential p-forms. As any p-form F induces
R an induced volume form on a p-dimensional
sub-manifold G defining so an integral G F . The exterior derivative d : Λp → Λp+1 satis-
fies d2 = 0 and defines an elliptic complex. There is a natural Hodge duality isomorphism
given called “Hodge star” ∗ : Λp → Λn−p . Given a p-form F ∈ Λp and a (p + 1)-dimensional
compact oriented sub-manifold G of M with boundary δG compatible with the orientation of
G, we have Stokes theorem:‘
R R
Theorem: hG, dF i = G dF = δG F = hδG, F i.
The theorem states that the exterior derivative d is dual to the boundary operator δ. If G
is a connected 1-manifold with boundary, it is a curve with boundary δG = {A, B}. A 1-
form can be integrated over the curve G by choosing the on G induced volume form r0 (t)dt
Rb
given by a curve parametrization [a, b] → G and integrate a F (r(t)) · r0 (t)dt, which is
the line integral. Stokes theorem is then the fundamental theorem of line integrals.
R b a 0-form0 f which is a scalar function the derivative df is the gradient F = ∇f . Then
Take
a
∇f (r(t)) · r (t) dt = f (B) − f (A). If G is a two dimensional surface with boundary δG and F
is a 1-form, then the 2-form dF is the curl of F . If G is given as a surface parametrization
r(u, v), one can applyR dF on the pair of tangent vectors ru , rv and integrate this dF (ru , rv ) over
the surface GR to get G dF . The Kelvin-Stokes theorem tells that this is the same than the
line integral δG F . In the case of M = R3 , where F = P dx + Qdy + Rdz can be identified with
a vector field FR R= [P, Q, R] and dF = ∇R ×R F and integration of a 2-form H over a parametrized
manifold G is R
H(r(u, v))(ru , rv ) = R
H(r(u, v)·ru ×rv dudv we get the classical Kelvin-
Stokes theorem. If F is a 2-form, then dF is a 3-form which can be integrated over a 3-
∗
manifold G. As d : Λ2 → Λ3 can via Hodge duality naturally 1
R R R be paired with d0R: RΛ → Λ ,
0
which is the divergence, the divergence theorem G

div(F ) dxdydz = δG
F · dS
relates a triple integral with a flux integral. Historical milestones start with the development
of the fundamental theorem of calculus (1666 Isaac Newton, 1668 James Gregory, Isaac
Barrow 1670 and Gottfried Leibniz 1693); the first rigorous proof was done by Cauchy in 1823
(the first textbook appearance in 1876 by Paul du Bois-Reymond). See [58]. In 1762, Joseph-
Louis Lagrange and in 1813 Karl-Friedrich Gauss look at special cases of divergence theorem,
Mikhail Ostogradsky in 1826 and George Green in 1828 cover the general case. Green’s theorem
in two dimensions was first stated by Augustin-Louis Cauchy in 1846 and Bernhard Riemann
in 1851. Stokes theorem first appeared in 1854 as an exam question but the theorem has
appeared already in a letter of William Thomson to Lord Kelvin in 1850, hence also the name
Kelvin-Stokes theorem. Vito Volterra in 1889 and Henri Poincaré in 1899 generalized the
theorems to higher dimensions. Differential forms were introduced in 1899 by Élie Cartan. The
50
OLIVER KNILL
d notation for exterior derivative was introduced in 1902 by Theodore de Donder. The ultimate
formulation above is from Cartan 1945. We followed Katz [235] who noticed that only in 1959,
this version has started to appear in textbooks.
115. Moments
The Hausdorff moment problem asks for R 1 necessary and sufficient conditions for a sequence
n
µn to be realizable as a moment sequence 0 x dµ(x) for a Borel probability measure on [0, 1].
One can Rstudy the problem also in higher dimensions: for a multi-index n = (n1 , . . . , nd ) denote
by µn = xn1 1 . . . xnd d dµ(x) the n’th moment of a signed Borel measure µ on the unit cube
I d = [0, 1]d ⊂ Rd . We say µn is a moment configuration if there exists a measure µ which
has µn as moments. If ei denotes the standard basis in Zd , define the partial
difference

k
Q ki k
Q n ki n Q d ni
(∆i a)n = an−ei − an and ∆ = i ∆i . We write n = i=1 ni and = i=1
Pn Pn1 Pnd k ki
and k=0 = k1 =0 · · · kd =0 . We say moments µn are Hausdorff bounded if there exists a

Pn n
constant C such that k=0 | (∆k µ)n | ≤ C for all n ∈ Nd . The theorem of Hausdorff-
k
Hildebrandt-Schoenberg is
Theorem: Hausdorff bounded moments µn are generated by a measure µ.
The above result is due to Theophil Henry Hildebrandt and Isaac Jacob Schoenberg from 1933.
[198]. Moments also allow to compare measures: a measure µ is called uniformly absolutely
continuous with respect to ν if there exists f ∈ L∞ (ν) such that µ = f ν. A positive probability
measure µ is uniformly absolutely continuous with respect to a second probability measure ν
if and only if there exists a constant C such that (∆k µ)n ≤ C · (∆k ν)n for all k, n ∈ Nd .
In particular it gives a generalization of a result of Felix Hausdorff from 1921 [191] assuring
that µ is positive if and only if (∆k µ)n ≥ 0 for all k, n ∈ Nd . An other special case is that
µ is uniformly
absolutely continuous with respect to Lebesgue measure ν on I d if and only if
n
|∆k µn | ≤ (n + 1)d for all k and n. Moments play an important role in statistics, when
k
looking at moment generating functions n µn tn of random variables X, where µn = E[X n ]
P
as well as in multivariate statistics, when looking at random vectors (X1 , . . . , Xd ), where
µn = E[X1n1 · · · Xdnd ] are multivariate moments. See [248]
116. Martingales
A sequence of random variables X1 , X2 , . . . on a probability space (Ω, A, P) is called a discrete
time stochastic process. We assume the Xk to be in L2 meaning that the expectation
E[Xk2 ] < ∞ for all k. Given a sub-σ algebra B of A, the conditional expectation E[X|B] is
the projection of L2 (Ω, A, P ) to L2 (Ω, B, P ). Extreme cases are E[X|A] = X and E[X|{∅, Ω}] =
E[X]. A finite set Y1 , . . . , Yn of random variables generates a sub- σ-algebra B of A, the
smallest σ-algebra for which all Yj are still measurable. Write E[X|Y1 , · · · , Yn ] = E[X|B],
where B is the σ-algebra generated by Y1 , · · · Yn . A discrete time stochastic process is called
a martingale if E[Xn+1 |X1 , · · · , Xn ] = E[Xn ] for all n. If the equal sign is replaced with ≤
Pn is called a super-martingale, if ≥ it is a sub-martingale.
then the process
2
The random
walk Xn = Y
k=1 k defined by a sequence of independent L random variables Yk is an
example of a martingale because independence implies E[Xn+1 |X1 , · · · , Xn ] = E[Xn+1 ] which
is E[Xn ] by the identical distribution assumption. If X and M are two discrete time stochastic
51
P the martingale transform (=discrete Ito integral) X · M as the process

processes, define
(X · M )n = nk=1 Xk (Mk − Mk−1 ). If the process X is bounded meaning that there exists a
constant C such that E[|Xk |] ≤ C for all k, then if M is a martingale, also X ·M is a martingale.
The Doob martingale convergence theorem is
Theorem: For a bounded super martingale X, then Xn converges in L1 .
The convergence theorem can be used to prove the optimal stopping time theorem which
tells that the expected value of a stopping time is the initial expected value. In finance it
is known as the fundamental theorem of asset pricing. If τ is a stopping time adapted
to a martingale Xk , it defines the random variable Xτ and E[Xτ ] = E[X0 ]. For a super-
martingale one has ≥ and for a sub-martingale ≤. The proof is obtained by defining the
Pmin(τ,n)−1
stopped process Xnτ = X0 + k=0 (Xk+1 − Xk ) which is a martingale transform and so
a martingale. The martingale convergence theorem gives a limiting random variable Xτ and
because E[Xnτ ] = E[X0 ] for all n, E[Xτ ] = E[X0 ]. This is rephrased as “you can not beat the
system” [427]. A trivial implication is that one can not for example design a strategy allowing
to win in a fair game by designing a “clever stopping time” like betting on “red” in roulette if
6 times “black” in a row has occurred. Or to follow the strategy to stop the game, if one has a
first positive total win, which one can always do by doubling the bet in case of losing a game.
Martingales were introduced by Paul Lévy in 1934, the name “martingale” (referring to the
just mentioned doubling betting strategy) was added in a 1939 probability book of Jean Ville.
The theory was developed by Joseph Leo Doob in his book of 1953. [120]. See [427].
117. Theorema Egregium

A Riemannian metric on a two-dimensional manifold S defines the quadratic form I = Edu2 +
2F dudv + Gdv 2 called first fundamental form on the surface. If r(u, v) is a parameterization
of S, then E = ru · ru , F = ru · rv and G = rv · rv . The second fundamental form of S
is II = Ldu2 + 2M dudv + N dv 2 , where L = ruu · n, M = ruv · n, N = rvv · n, written using
the normal vector n = (ru × rv )/|ru × rv |. The Gaussian curvature K = det(II)/det(I) =
(LN − M 2 )/(EG − F 2 ). depends on the embedding r : R → S in space R3 , but it actually only
depends on the intrinsic metric, the first fundamental form. This is the Theorema egregium
of Gauss:
Theorem: The Gaussian curvature only depends on the Riemannian metric.
Gauss himself already gave explicit formulas, but a formula of Brioschi gives the curvature K
explicitly as a ratio of determinants involving E, F, G as well as and first and second derivatives
of them. In the case when the surface is given as a graph z = f (x, y), one can give K =
D/(1 + |∇f |2 )2 , where D = (fxx fyy − fxy2
) is the discriminant and (1 + |∇f |2 )2 = det(II). If
the surface is rotated in space so that (u, v) is a critical point for f , then the discriminant
D is equal to the curvature. One can see the independence of the embedding also from the
Puiseux formula K = 3(|S0 (r)| − S(r))/(πr3 ), where |S0 (r)| = 2πr is the circumference of
the circle S0 (r) in the flat case and |S(r)| is the circumference of the geodesic circle of radius
r on S. The theorem Egregium also follows from Gauss-Bonnet as the later allows to write the
curvature in terms of the angle sum of a geodesic infinitesimal triangle with the angle sum π
of a flat triangle. As the angle sums are entirely defined intrinsically, the curvature is intrinsic.
The Theorema Egregium was found by Karl-Friedrich Gauss in 1827 and published in 1828
52
OLIVER KNILL
in “Disquisitiones generales circa superficies curvas”. It is not an accident, that Gauss was
occupied with concrete geodesic triangulation problems too.
118. Entropy
Given a random variable X on a probability space (Ω, A, P) which is finite and discrete
P in the
sense that it takes only finitely many values, the entropy is defined as S(X) = − x px log(px ),
where px = P[X = x]. To compare, for a random variable X with cumulative distribution
0
R = P[X ≤ x] having a continuous derivative F = f , the entropy is defined as
function F (x)
S(X) = − f (x) log(f (x)) dx, allowing the value −∞ if the integral does not converge. (We
always read p log(p) = 0 if p = 0.) In the continuous case, one also calls this the differential
entropy. Two discrete random variables X, Y are called independent if one can realize them
on a product probability space Ω = A × B so that X(a, b) = X(a) and Y (a, b) = Y (b) for
some functions X : A → R, Y : B → R. Independence implies that the random variables are
uncorrelated, E[XY ] = E[X]E[Y ] and that the entropy adds up S(XY ) = S(X) + S(Y ).
We can write S(X) = E[log(W (x))], where W is the “Wahrscheinlichkeit” random variable
assigning to ω ∈ Ω the value W (ω) = 1/px if X(ω) = x. P Let us say, a functional on discrete
random variables is additive if it is of the form H(X) = x f (px ) for some continuous function
f for which f (t)/t is monotone. We say it is multiplicative if H(XY ) = H(X) + H(Y ) for
independent random variables. The functional is normalized if H(X) = log(4) if X is a
random variable taking two values {0, 1} with probability p0 = p1 = 1/2. Shannon’s theorem
is:
Theorem: Any normalized, additive and multiplicative H is entropy S.
The word “entropy” was introduced by Rudolf Clausius in 1850 [349]. Ludwig Bolzmann saw
the importance of dtd S ≥ 0 in the context of heat and wrote in 1872 S = kB log(W ), where
W (x) = 1/px is the inverse “Wahrscheinlichkeit” thatPa state has the value x. His equation
is understood as the expectation S = kB E[log(W )] = x px log(W (x)) which is the Shannon
entropy, introduced in 1948 by Claude Shannon in the context of information theory. (Shannon
characterized functionals H with the property that if H is continuous in p, then for random
variables Hn with px (Hn ) = 1/n, one has H(Xn )/n ≤ H(Xm )/m if n ≤ m and if X, Y are
two random variables so that theP finite σ-algebras A defined by X is a sub-σ-algebra B defined
by Y , then H(Y ) = H(X) + x px H(Yx ), where Yx (ω) = Y (ω) for ω ∈ {X = x}. One can
show that these Shannon conditions are equivalent to the combination of being additive and
multiplicative. In statistical thermodynamics, where px is the probability of a micro-state,
then kB S is also called the Gibbs entropy, where kB is the Boltzmann constant. For
general random variables X on (Ω, A, P) and a finite σ-sub-algebra B, Gibbs looked in 1902 at
course grained entropy, which is the entropy of the conditional expectation Y = E[X|B|,
which is now a random variable Y taking only finitely many values so that entropy is defined.
See [363].
119. Mountain Pass

Let H be a Hilbert space, and let f be a twice Fréchet differentiable function from H to
R. The Fréchet derivative A = f 0 at a point x ∈ H is a linear operator A satisfying
f (x + h) − f (x) − Ah = o(h) for all h → 0. A point x ∈ H is called a critical point of f
if f 0 (x) = 0. The functional satisfies the Palais-Smale condition, if every sequence xk in H
for which {f (xk )} is bounded and f 0 (xk ) → 0, has a convergent subsequence in the closure of
53
{xk }k∈N . A pair of points a, b ∈ H defines a mountain pass, if there exist > 0 and r > 0
such that f (x) ≥ f (a) + on Sr (a) = {x ∈ H | ||x − a|| = r}, f is not constant on Sr (a) and
f (b) ≤ f (a). A critical point is called a saddle if it is neither a maximum nor a minimum of
f.
Theorem: If a Palais-Smale f has a mountain pass, it features a saddle.
The idea is to look at all continuous paths γ from a to b parametrized by t ∈ [0, 1]. For each
path γ, the value cγ = f (γ(t)) has to be maximal for some time t ∈ [0, 1]. The infimum over
all these critical values cγ is a critical value of f . The mountain pass condition leads to a
“mountain ridge” and the critical point is a “mountain pass”, hence the name. The example
(2 exp(−x2 − y 2 ) − 1)(x2 + y 2 ) with a = (0, 0), b = (1, 0) shows that the non-constant condition
is necessary for a saddle point on Sr (a) with r = 1/2. The reason for sticking with a Hilbert
space is that it is easier to realize the compactness condition due to weak star compactness of
the unit ball. But it is possible to weaken the conditions and work with a Banach manifolds X
continuous Gâteaux derivatives: f 0 : X → X ∗ if X has the strong and X ∗ the weak-∗ topology.
It is difficult to pinpoint historically the first use of the mountain pass principle as it must have
been known intuitively since antiquity. The crucial Palais-Smale compactness condition
which makes the theorem work in infinite dimensions appeared in 1964. [24] calls it condition
(C), a notion which already appeared in the original paper [324].
120. Exponential sums

Given a smoothPfunction f : R → R which maps integers to integers, P one can look at expo-
b n−1
nential sums x=a exp(iπf (x)) An example is the Gaussian sum x=0 exp(iαx2 ). There
are lots of interesting relations and estimates. One P
of the magical formulas is the Landsberg-
Schaar relations for the finite sums S(q, p) = √1p p−1 2
x=0 exp(iπx q/p).
Theorem: If p, q are positive and odd integers, then S(2q, p) = eiπ/4 S(−p, 2q).
√ P
One has S(1, p) = (1/ p) p−1 2
x=0 exp(ix /p) = 1 for all positive integers p and S(2, p) =
√
(eiπ/4 / p) p−1 2
P
x=0 exp(2ix /p) = 1 if p = 4k + 1 and i if p = 4k − 1. The method of expo-
nential sums has been expanded especially by Vinogradov’s papers [413] and used for number
theory like for quadratic reciprocity [311]. The topic is of interest also outside of number the-
ory. Like in dynamical systems theory as Fürstenberg has demonstrated. An ergodic theorist
would look at the dynamical system T (x, y) = (x + 2y + 1, y + 1) on the 2-torus T2 = R2 /(πZ)2
and define gα (x, y) = exp(iπxα). Since the orbit of P this toral map is T n (1, 1) = (n2 , n), the
exponential sum can be written as a Birkhoff sum p−1 k
k=0 gq/p (T (1, 1)) which is a particular
orbit of a dynamical system. Results as those mentioned above show that the random walk
√
grows like p, similarly as in a random setting. Now, since the dynamical system is minimal,
the growth rate should not depend on the initial point and πq/p should be replaceable by any
irrational α and no more be linked to the length of the orbit. The problem is then to study
Pt−1
the growth rate of the stochastic process S (x, y) = k=0 g(T k (x, y)) (= sequence of random
t
variables) for any continuous g with zero expectation which by Fourier boils down to look at
exponential sums. Of course S t (x, y)/t → 0 by Birkhoff’s ergodic theorem, but as in the law
of iterated logarithm one is interested in precise growth rates. This can be subtle. Already
in the simpler case of an integrable T (x) = x + α on the 1-torus, there is Denjoy-Koskma
theory which shows that the growth rate depends on Diophantine properties of πα. Unlike for
54
OLIVER KNILL
irrational rotations, the Fürstenberg type skew systems T leading to the theta functions are not
integrable: it is not conjugated to a group translation (there is some randomness, even-so weak
as Kolmogorov-Sinai entropy is zero). The dichotomy between structure and randomness and
especially the similarities between dynamical and number theoretical set-ups has been discussed
in [400].
121. Sphere theorem

A compact Riemannian manifold M is said to have positive curvature, if all sectional
curvatures are positive. The sectional curvature at a point x ∈ M in the direction of the 2-
dimensional plane Σ ⊂ Tx M is defined as the Gaussian curvature of the surface expx (Σ) ⊂ M at
the point. In terms of the Riemannian curvature tensor R : Tx M 4 → R and an orthonormal
basis {u, v} spanning Σ, this is R(u, v, u, v). The curvature is called quarter pinched, if it
the sectional curvature is in the interval (1, 4] at all points x ∈ M . In particular, a quarter
pinched manifold is a manifold with positive curvature. We say here, a compact Riemannian
manifold is a sphere if it is homeomorphic to a sphere. The sphere theorem is:
Theorem: A simply-connected quarter pinched manifold is a sphere
The theorem was proven by Marcel Berger and Wilhelm Klingenberg in 1960. That a pinching
condition would imply a manifold to be a sphere had been conjectured already by Heinz Hopf.
Hopf himself proved in 1926 that constant sectional curvature implies that M is even isometric
to a sphere. Harry Rauch, after visiting Hopf in Zürich in the 1940’s proved that a 3/4-
pinched simply connected manifold is a sphere. In 2007, Simon Brendle and Richard Schoen
proved that the theorem even holds if the statement M is a d-sphere (meaning that M is
diffeomorphic to the Euclidean d-sphere {|x|2 = 1} ⊂ Rd+1 ). This is the differentiable sphere
theorem. Since John Milnor had given in 1956 examples of spheres which are homeomorphic
but not diffeomorphic to the standard sphere (so called exotic spheres, spheres which carry
a smooth maximal atlas different from the standard one), the differentiable sphere theorem
is a substantial improvement on the topological sphere theorem. It needed completely new
techniques, especially the Ricci flow ġ = −2Ric(g) of Richard Hamilton which is a weakly
parabolic partial differential equation deforming the metric g and uses the Ricci curvature
Ric of g. See [40, 57].
122. Word problem

The word problem in a finitely presented group G = (g|r) with generators g and re-
lations r is the problem to decide, whether a given set of two words v, w represent the same
group element in G or not. The word problem is not solvable in general. There are concrete
finitely presented groups in which it is not. The following theorem of Boone and Higman relates
the solvability to algebra. A group is simple if its only normal subgroup is either the trivial
group or then the group itself.
Theorem: Finitely presented simple groups have a solvable word problem.
More generally, if G ⊂ H ⊂ K where H is simple and K is finitely presented, then G has

a solvable word problem. Max Dehn proposed the word problem in 1911. Pyotr Novikov in
1955 proved that the word problem is undecidable for finitely presented groups. William W.
Boone and Graham Higman proved the theorem in 1974 [47]. Higman would in the same year
55
also find an example of an infinite finitely presented simple group. The non-solvability of the
word problem implies the non-solvability of the homeomorphism problem for n-manifolds with
n ≥ 4. See [435].
123. Finite simple groups

A finite group (G, ∗, 1) is a finite set G with an operation ∗ : G×G → G and 1 element, such
that the operation is associative (a ∗ b) ∗ c = a ∗ (b ∗ c), for all a, b, c, such that a ∗ 1 = 1 ∗ a = a
for every a and such that every a has an inverse a−1 satisfying a ∗ a−1 = 1. A group G is simple
if the only normal subgroups of G are the trivial group {1} or the group itself. A subgroup
H of G is called normal if gH = Hg for all g. Simple groups play the role of the primes
in the integers. A theorem of Jordan-Hölder is that a decomposition of G into simple groups
is essentially unique up to permutations and isomorphisms. The classification theorem of
finite simple groups is
Theorem: Every finite simple group is cyclic, alternating, Lie or sporadic.
There are 18 so called regular families of finite simple groups made of cyclic, alternating
and 16 Lie type groups. Then there are 26 so called sporadic groups, in which 20 are happy
groups as they are subgroups or sub-quotients of the monster and 6 are pariahs, outcasts
which are not under the spell of the monster. The classification was a huge collaborative effort
with more than 100 authors covering 500 journal articles. According to Daniel Gorenstein, the
classification was completed 1981 and fixes were applied until 2004 (Michael Aschbacher and
Stephen Smith resolving the last problems which lasted several years) leading to a full proof of
1300 pages. A second generation cleaned-out proof written with more details is under way and
currently has 5000 pages. Some history is given in [377].
124. God number

Given a finite finitely presented group G = (g|r) like for example the Rubik group. It defines
the Cayley graph Γ in which the group elements are the nodes and where two nodes a, b
are connected if there is a generator x in in g such that xa = b. The diameter of a graph
is the largest geodesic distance between two nodes in Γ. It is also called the God number
of the puzzle. The Rubik cube is an example of a finitely presented group. The original
3 × 3 × 3 cube allows to permute the 26 boundary cubes using the 18 possible rotations of the 6
faces as generators. From the X = 8!12!38 212 possible ways to physically build the cube, only
|G| = X/12 = 43252003274489856000 are present in the Rubik group G. Some of the positions
“quarks” [165] can not be realized but combinations of them “mesons” or “baryons” can.
Theorem: The God number of the Rubik cube is 20.
This means that from any position, one could, in principle solve the puzzle in 20 moves. Note
that one has to specify clearly the generators of the group as this defines the Cayley graph and
so a metric on the group. The lower bound 18 had already been known in 1980 as counting
the possible moves with 17 moves produces less elements. The lower bound 20 came in 1995
when Michael Reid proved that the superflip position (where the edges are all flipped but
corners are correct) needs 20 moves. In July 2010, using about 35 CPU years, a team around
Tomas Rokicki established that the God number is 20. They partitioned the possible group
positions into roughly 2 billion sets of 20 billions positions each. Using symmetry they reduced
it to 55 million positions, then found solutions for any of the positions in these sets. [137]
56
OLIVER KNILL
It appears silly to put a God number computation as a fundamental theorem, but the status
of the Rubik cube is enormous as it has been one of the most popular puzzles for decades
and is a prototype for many other similar puzzles, the choice can be defended. 1 One can
ask to compute the god number of any finitely presented finite group. Interesting in general
is the complexity of evaluating that functional. The simplest nontrivial Rubik cuboid is
the 2 × 2 × 1 one. It has 6 positions and 2 generators a, b. The finitely presented group
is {a, b|a2 = b2 = (ab)3 = 1} which is the dihedral group D3 . Its group elements are
G = {1, a = babab, ab = baba, aba = bab, abab = ba, ababa = b}. The group is isomorphic to the
symmetry group of the equilateral triangle, generated by the two reflections a, b at two
altitude lines. The God number of that group is 3 because the Cayley graph Γ is the cyclic
graph C6 . The puzzle solver has here “no other choice than solving the puzzle” than to make
non-trivial move in each step. See [224] or [33] for general combinatorial group theory.
125. Sard Theorem

Let f : M → N be a smooth map between smooth manifolds M, N of dimension dim(M ) = m
and dim(N ) = n. A point x ∈ M is called a critical point of f , if the Jacobian n × m matrix
df (x) has rank both smaller than m and n. If C is the set of critical points, then f (C) ⊂ N
is called the critical set of f . The volume measure on N is a choice of a volume form,
obtained for example after introducing a Riemannian metric. Sard’s theorem is
Theorem: The critical set of f : M → N has zero volume measure in N .
The theorem applied to smooth map f : M → R tells that for almost all c, the set f −1 (c)
is a smooth hypersurface of M or then empty. The later can happen if f is constant. We
assumed C ∞ but one can relax the smoothness assumption of f . If n ≥ m, then f needs only
to be continuously differentiable. If n < m, then f needs to be in C m−n+1 . The case when N
is one-dimensional has been covered by Antony Morse (who is unrelated to Marston Morse)
in 1939 and by Arthur Sard in general in 1942. A bit confusing is that Marston Morse (not
Antony) covered the case m = 1, 2, 3 and Sard in the case m = 4, 5, 6 in unpublished papers
before as mentioned in a footnote to [357]. Sard also notes already that examples of Hassler
Whitney show that the smoothness condition can not be relaxed. Sard formulated the results
for M = Rm and N = Rn (by the way with the same choice f : M → N as done here and not
as in many other places). The manifold case appears for example in [386].
126. Elliptic curves

An elliptic curve is a plane algebraic curve defined by the points satisfying the Weierstrass
equation y 2 = x3 + ax + b = f (x). One assumes the curve to be non-singular, meaning
that the discriminant ∆ = −16(4a3 + 27b2 ) is not zero. This assures that there are no
cusps nor multiple roots for the simple reason that the explicit solution formulas for roots of
f (x) = 0 involves only square roots of ∆. A curve is an Abelian variety, if it carries an
Abelian algebraic group structure, meaning that the addition of a point defines a morphism of
the variety.
1I presented the God number problem in the 80ies as an undergraduate in a logic seminar of Ernst Specker
and the choice of topic had been objected to by Specker himself as a too “narrow problem”. But for me, the
Rubik cube and its group theoretical properties have “cult status” and was one of the triggers to study math.
57
Theorem: Elliptic curves are Abelian varieties.
The theorem seems first have been realized by Henry Poincaré in 1901. Weierstrass before had
used the Weierstrass P function earlier in the case of elliptic curves over the complex plane. To
define the group multiplication, one uses the chord-tangent construction: first add point O
called the point at infinity which serves as the zero in the group. Then define −P as the point
obtained by reflecting at the x-axes. The group multiplication between two different points
P, Q on the curve is defined to be −R if R is the point of intersection of the line through P, Q
with the curve. If P = Q, then R is defined to be the intersection of the tangent with the curve.
If there is no intersection that is if P = Q is an inflection point, then one defines P + P = −P .
Finally, define P + O = O + P = P and P + (−P ) = 0. This recipe can be explicitly given
in coordinates allowing to define the multiplication in any field of characteristic different from
2 or 3. The group structure on elliptic curves over finite fields provides a rich source of finite
Abelian groups which can be used for cryptological purposes, the so called elliptic curve
cryptograph ECC. Any procedure, like public key, Diffie-Hellman or factorization attacks on
integers can be done using groups given by elliptic curves. [415].
127. Billiards
Billiards are the geodesic flow on a smooth compact n-manifold M with boundary. The dy-
namics is extended through the boundary by applying the law of reflection. While the flow
of the geodesic X t is Hamiltonian on the unit tangent bundle SM , the billiard flow is only
piecewise smooth and also the return map to the boundary is not continuous in general but it
is a map preserving a natural volume so that one can look at ergodic theory. Already difficult
are flat 2-manifolds M homeomorphic to a disc having convex boundary homeomorphic to a
circle. For smooth convex tables this leads to a return map T on the annulus X = T × [−1, 1]
which is C r−1 smooth if the boundary is C r [121]. It defines a monotone twist map: in the
sense that it preserves the boundary, is area and orientation preserving and satisfies the twist
condition that y → T (x, y) is strictly monotone. A Bunimovich stadium is the 2-manifold
with boundary obtained by taking the convex hull of two discs of equal radius in R with dif-
ferent center. The billiard map is called chaotic, if it is ergodic and the Kolmogorov-Sinai
entropy is positive. By Pesin theory, this metric entropy is the Lyapunov exponent which
is the exponential growth rate of the Jacobian dT n (and constant almost everywhere due to
ergodicity). There are coordinates in the tangent bundle of the annulus X in which dT is the
composition of a horizontal shear with strength L(x, y), where L is the trajectory length before
the impact with a vertical shear with strength −2κ/ sin(θ) where κ(x) is the curvature of the
curve at the impact x and y = cos(θ), with impact angle θ ∈ [0, π] between the tangent and
the trajectory.
Theorem: The Bunimovich stadium billiard is chaotic.

Jacques Hadmard in 1898 and Emile Artin in 1924 already looked at the geodesic flow on
a surface of constant negative curvature. Yakov Sinai constructed in 1970 the first chaotic
billiards, the Lorentz gas or Sinai billiard. An example, where Sinai’s result applies is the
hypocycloid x1/3 + y 1/3 = 1. The Bernoulli property was established by Giovanni Gallavotti
and Donald Ornstein in 1974. In 1973, Vladimir Lazutkin proved that a generic smooth convex
two-dimensional billiard can not be ergodic due to the presence of KAM whisper galleries
using Moser’s twist map theorem. These galleries are absent in the presence of flat points (by a
58
OLIVER KNILL
theorem of John Mather) or points, where the curvature is unbounded (by a theorem of Andrea
Hubacher). Leonid Bunimovich constructed in 1979 the first convex chaotic billiard. No smooth
convex billiard table with positive Kolmogorov-Sinai entropy is known. A candidate is the real
analytic x4 + y 4 = 1. Various generalizations have been considered like in [430]. A detailed
proof that the Bunimovich stadium is measure theoretically conjugated to a Bernoulli system
(the shift on a product space) is surprisingly difficult: one has to show positive Lyapunov
exponents on a set of positive measure. Applying Pesin theory with singularities (Katok-
Strelcyn theory) gives a Markov process. One needs then to establish ergodicity using a method
of Eberhard Hopf of 1936 which requires to understand stable and unstable manifolds [82]. See
[398, 414, 308, 164, 230, 82] for sources on billiards.
128. Uniformization
A Riemann surface is a one-dimensional complex manifold. This means is is a connected
two dimensional manifold so that the transition functions o the atlas are holomorphic mappings
of the complex plane. It is simply connected if its fundamental group is trivial meaning that
its genus b1 is zero. Two Riemann surfaces are conformally equivalent or simply equivalent
if they are equivalent as complex manifolds, that is if a bijective morphism f between them
exists. A map f : S → S 0 is holmorphic if for every choice of coordinates φ : S → C and
ψ 0 : S 0 → C, the maps φ0 ◦ f ◦ φ−1 is holomorphic. The curvature is the Gaussian curvature
of the surface. The uniformization theorem is:
Theorem: A Riemann surface is equivalent to one with constant curvature.
This is a “geometrization statement” and means that the universal cover of every Riemann sur-
face is conformally equivalent to either a Riemann sphere (positive curvature), a complex
plane (zero curvature) or a unit disk (negative curvature). It implies that any region G ⊂ C
whose complement contains 2 or more points has a universal cover which is the disk which
especially implies the Riemann mapping theorem assuring that and region U homeomor-
phic to a disk is conformally equivalent to the unit disk. (see [74]. For a detailed treatement
of compact Riemann surfaces, see [161]. It also follows that all Riemann surfaces (without
restriction of genus) can be obtained as quotients of these three spaces: for the sphere one
does not have to take any quotient, the genus 1 surfaces = elliptic curves can be obtained as
quotients of the complex plane and any genus g > 1 surface can be obtained as quotients of the
unit disk. Since every closed 2-dimensional orientable surface is characterized by their genus g,
the uniformization theorem implies that any such surface admits a metric of constant curva-
ture. Teichmüller theory parametrizes the possible metrics, and there are 3g − 3 dimensional
parameters for g ≥ 2, whereas for g = 0 there is one and for g = 1 a moduli space H/SL2 (Z).
In higher dimensions, close to the uniformization theorem comes the Killing-Hopf theorem
telling that every connected complete Riemannian manifold of constant sectional curvature
and dimension n is isometric to the quotient of a sphere Sn , Euclidean space Rn or Hyperbolic
n-space Hn . Constant curvature geometry is either Elliptic, Parabolic=Euclidean or Hyperbolic
geometry. Complex analysis has rich applications in complex dynamics [34, 300, 74] and relates
to much more geometry [297].
129. Control Theory

A Kalman filter is an optional estimates algorithm of a linear dynamic system from a series
of possibly noisy measurements. The idea is similar as in a dynamic Bayesian network or
59
hidden Markov model. It applies both to differential equations ẋ(t) = Ax(t) + Bu(t) +
Gz(t) as well as discrete dynamical system x(t + 1) = Ax(t) + Bu(t) + Gz(t), where u(t)
is external input and z(t) input noise given by independent identically distributed usually
Gaussian random variables. Kalman calls this a Wiener problem. One does not see the
state x(t) of the system but some output y(t) = Cx(t) + Du(t). The filter then “filters out”
or “learns” the best estimate x∗ (t) from the observed data y(t). The linear space X is defined
as the vector space spanned by the already observed vectors. The optimal solution is given by
a sophisticated dynamical data fitting.
Theorem: The optimal estimate x∗ is the projection of y onto X.
This is the informal 1-sentence description which can be found already in Kalman’s article.
Kalman then gives explicit formulas which generate from the stochastic difference equation
a concrete deterministic linear system. For a modern exposition, see [286]. This is the
Kalman filter. It is named after Rudolf Kalman who wrote [226] in 1960. Kalman’s paper
is one of the most cited papers in applied mathematics. The ideas were used in the Apollo
and Space Shuttle program. Similar ideas have been introduced in statistics by the Danish
astronomer Thorvald Thiele and the radar theoretician Peter Swerling. There are also nonlinear
version of the Kalman filter which is used in nonlinear state estimation like navigation systems
and GPS. The nonlinear version uses a multi-variate Taylor series expansion to linearize about
a working point. See [138, 286].
130. Zariski main theorem

A variety is called normal if it can be covered by open affine varieties whose rings of functions
are normal. A commutative ring is called normal if it has no non-zero nilpotent elements and
is integrally closed in its complete ring of fractions. For a curve, a one-dimensional variety,
normality is equivalent to being non-singular but in higher dimensions, a normal variety still
can have singularities. The normal complex variety is called unibranch at a point x ∈ X if
there are arbitrary small neighborhoods U of x such that the set of non-singular points of U is
connected. Zariski’s main theorem can be stated as:
Theorem: Any closed point of a normal complex variety is unibranch.
Oscar Zariski proved the theorem in 1943. To cite [309], “it was the final result in a foundational
analysis of birational maps between varieties. The ’main Theorem’ asserts in a strong sense
that the normalization (the integral closure) of a variety X is the maximal variety X 0 birational
over X, such that the fibres of the map X 0 → X are finite. A generalization of this fact
became Alexandre Grothendieck’s concept of the ’Stein factorization’ of a map. The result
has been generalized to schemes X, which is called unibranch at a point x if the local ring
at x is unibranch. A generalization is the Zariski connectedness theorem from 1957: if
f : X → Y is a birational projective morphism between Noetherian integral schemes, then the
inverse image of every normal point of Y is connected. Put more colloquially, the fibres of a
birational morphism from a projective variety X to a normal variety Y are connected. It implies
that a birational morphism f : X → Y of algebraic varieties X, Y is an open embedding into a
neighbourhood of a normal point y if f −1 (y) is a finite set. Especially, a birational morphism
between normal varieties which is bijective near points is an isomorphism. [189, 309]
60
OLIVER KNILL
131. Poincaré’s last theorem

A homeomorphism T of an annulus X = T × [0, 1] is measure preserving if it preserves the
Lebesgue (area) measure and preserves the orientation of X. As a homeomorphism it induces
also homeomorphisms on each of the two boundary circles. It is called twist homeomorphism,
if it rotates the boundaries into different directions.
Theorem: A twist map on an annulus has at least two fixed points.
This is called the Poincaré-Birkhoff theorem or Poincaré’s last theorem. It was stated by
Henry Poincaré in 1912 in the context of the three body problem. Poincaré already gave an
index argument for the existence of one fixed point gives a second. The existence of the first was
proven by George Birkhoff in 1913 and in 1925, he added the precise argument for the existence
of the second. The twist condition is necessary as the rotation of the annulus (r, θ) → (r, θ + 1)
has no fixed point. Also area preservation is necessary as (r, θ) → (r(2 − r), θ + 2r − 1) shows.
[45, 63]
132. Geometrization
A closed manifold M is a smooth compact manifold without boundary. A closed manifold
is simply connected if it is connected and the fundamental group is trivial meaning that
every closed loop in M can be pulled together to a point within M : (if r : T → M is a
parametrization of a closed path in M , then there exists a continuous map R : T × [0, 1] → M
such that R(0, t) = r(t) and R(1, t) = r(0). We say that M is 3-sphere if M is homeomorphic
to a 3-dimensional sphere {(x1 , x2 , x3 , x4 ) ∈ R4 | x21 + x22 + x23 + x23 = 1}.
Theorem: A closed simply connected 3-manifold is a 3-sphere.
Henry Poincaré conjectured this in 1904. It remained the Poincaré conjecture until its
proof of Grigori Perelman in 2006 [305]. In higher dimensions the statement was known as
the generalized Poincaré conjecture, the case n > 4 had been proven by Stephen Smale
in 1961 and the case n = 4 by Michael Freedman in 1982. A d-homotopy sphere is a closed
d-manifold that is homotopic to a d-sphere. (A manifold M is homotopic to a manifold N if
there exists a continuous map f : M → N and a continuous map g : N → M such that the
composition g ◦ f : M → M is homotopic to the identity map on M (meaning that there exists
a continuous map F : M × [0, 1] → M such that F (x, 0) = g(f (x)) and F (x, 1) = x) and the
map f ◦ g : N → N is homotopic to the identity on N .) The Poincaré conjecture itself, the case
d = 3, was proven using a theory built by Richard Hamilton who suggested to use the Ricci
flow to solve the conjecture and more generally the geometrization conjecture of William
Thurston: every closed 3-manifold can be decomposed into prime manifolds which are of 8
˜
types, the so called Thurston geometries S 3 , E 3 , H 3 , S 2 × R, H 2 × R, SL(2, R), Nil, Solv. If
the statement M is a sphere is replaced by M is diffeomorphic to a sphere, one has the
smooth Poincaré conjecture. Perelman’s proof verifies this also in dimension d = 3. The
smooth Poincaré conjecture is false in dimension d ≥ 7 as d-spheres then admit non-standard
smooth structures, so called exotic spheres constructed first by John Milnor. For d = 5 it is
true following result of Dennis Barden from 1964. It is also true for d = 6. For d = 4, it is open,
and called “the last man standing among all great problems of classical geometric topology”
[282]. See [306] for details on Perelman’s proof.
61
133. Steinitz theorem

A non-empty finite simple connected graph G is called planar if it can be embedded in the
plane R2 without self crossings. The abstract edges of the graph are then realized as actual
curves in the plane connecting two vertices. The embedding of G in the plane subdivides the
plane into a finite collection F of simply connected regions called faces. Let v = |V | is the
number of vertices, e = |E| the number of edges and f = |F | is the number of faces. A planar
graph is called polyhedral if it can be realized as a convex polyhedron, a convex hull of
finitely many points in R3 . A graph is called 3-connected, if it remains connected also after
removing one or two of its vertices. A connected, planar 3-connected graph is also called a
3-polyhedral graph. The Polyhedral formula of Euler combined with Steinitz’s theorem
means:
Theorem: G planar ⇒ v − e + f = 2. Planar 3-connected ⇔ polyhedral.
The Euler polyhedron formula which was first seen in examples by René Descartes [4] but seen
by Euler in 1750 to work for general planar graphs. Euler already gave an induction proof
(also in 1752) but the first complete proof appears having been given first by Legendre in 1794.
The Steinitz theorem was proven by Ernst Steinitz in 1922, even so he obtained the result
already in 1916. In general, a planar graph always defines a finite generalized CW complex in
which the faces are the 2-cells, the edges are the 1-cells and the vertices are the 0-cells. The
embedding in the plane defines then a geometric realization of this combinatorial structure
as a topological 2-sphere. But the realization is not necessarily be achievable in the form of a
convex polyhedron. Take a tree graph for example, a connected graph without triangles and
without closed loops. It is planar but it is not even 2-connected. The number of vertices v and
the number of edges e satisfy v − e = 1. After embedding the tree in the plane, we have one
face so that f = 1. The Euler polyhedron formula v − e + f = 2 is verified but the graph is not
polyhedral. Even in the extreme case where G is a one-point graph, the Euler formula holds:
in that case there is v = 1 vertex, e = 0 edges and f = 1 faces (the complement of the point
in the plane) and still v − e + f = 2. The 3-connectedness assures that the realization can be
done using convex polyhedra. It is even possible to have force the vertices of the polyhedron
to be on the integer lattice. See [176, 438]. In [176], it is stated that the Steinitz theorem is
“the most important and deepest known result for 3-polytopes”.
134. Hilbert-Einstein action

Let (M, g) be a smooth 4-dmensional Lorentzian manifold which is symptotically flat. A
Lorentzian manifold is a 4 dimensional pseudo Riemannian manifold with signature (1, 3) which
in the flat case is dx2 + dy 2 + dz 2 − dt2 . We assume that the volume form dµ has the property
that the scalar curvature R is in L1 (M, R dµ). One can now look at the variational problem
to find extrema of the functional g → M RRdµ. More generally, one can add a Lagrangian
L one uses the Hilbert-Einstein functional M R/κ + Ldµ, where κ = 8πG/c4 is the Einstein
constant. Let Rij be the Ricci tensor, a symmetric tensor, and Tij the energy-momentum
tensor. The Einstein field equations are
Theorem: Gij = Rij − gij R/2 = κTij
These are the Euler-Lagrange equation of an infinite dimensional extremization problem. The
variational problem was proposed by David Hilbert in 1915. Einstein published in the same
62
OLIVER KNILL
year the general theory of relativity. In the case of a vacuum: T = 0, solutions g define
Einstein manifolds (M, g). An example of a solution to the vaccuum Einstein equations
different from the flat space solution is the Schwarzschild solution, which was found also in
1915 and published in 1916. It is the metric given in spherical coordinates as −(1 − r/ρ)c2 dt2 +
(1−r/ρ)−1 dρ2 +ρ2 dφ2 +ρ2 sin2 φdθ2 , where r is the Schwarzschild radius, ρ the distance to the
singularity, θ, φ are the standard Euler angles (longitude and colatitude) in calculus. The
metric solves the Einstein equations for ρ > r. The flat metric −c2 dt2 +dρ2 +ρ2 dθ2 +ρ2 sin2 θdφ2
describes the vacuum and the Schwarzschild solution describes the gravitational field near a
massive body. Intuitively, the metric tensor g is determined by g(v, v), and the Ricci tensor
by R(v, v) which is 3 times the average sectional curvature over all planes passing through
a plane through v. The scalar curvature is 6 times the average over all sectional curvatures
passing through a point. See [104, 85].
135. Hall stable marriage

Let X be a finite set and A be a family of finite subset A of X. A transversal of A is an
injective function f : A → X such that f (A) ∈ A for all A ∈ A. TheSset A satisfies the
marriage condition if for every finite subset B of A, one has |B| ≤ | A∈B A|. The Hall
marriage theorem is
Theorem: A has a transversal ⇔ A satisfies marriage condition.
The theorem was proven by Philip Hall in 1935. It implies for example that if a deck of cards
with 52 cards is partitioned into 13 equal sized piles, one can chose from each deck a card so
that the 13 cards have exactly one card of each rank. The theorem can be deduced from a
result in graph geometry: if G = (V, E) = (X, ∅) + (Y, ∅) is a bipartite graph, then a matching
in G is a collection of edges which pairwise have no common vertex. For a subset W of X, let
S(W ) denote the set of all vertices adjacent to some element in W . The theorem assures that
there is an X-saturating matching (a matching covering X) if and only if |W | ≤ |S(W )| for
every W ⊂ X. The reason for the name “marriage” is the situation that X is a set of men and
Y a set of women and that all men are eager to marry. Let Ai be the set of women which could
make a spouse for the i’th man, then marrying everybody off is an X-saturating matching. The
condition is that any set of k men has a combined list of at least k women who would make
suitable spouses. See [64].
136. Mandelbulb
The Mandelbrot set M = M2,2 is the set of vectors c ∈ R2 for which T (x) = x2 + c leads
to a bounded orbit starting at 0 = (0, 0), where x2 has the polar coordinates (r2 , 2θ) if x
has the polar coordinates (r, θ). This construction can be done in higher dimensions too: The
Mandelbulb set M3,8 is defined as the set of vectors c ∈ R3 for which T (x) = x8 + c leads to
a bounded orbit starting at 0 = (0, 0, 0), where x8 has the spherical coordinates (ρ8 , 8φ, 8θ)
if x has the spherical coordinates (ρ, φ, θ). Like the Mandelbrot set it is a compact set. The
topology of M8 is unexplored. Also like in the complex plane, one could look at the dynamics
of a polynomials p = a0 + a1 x + · · · + ar xr in Rn . If (ρ, φ1 , . . . , φn−1 ) are spherical coordinates,
then x → xm = (ρm , mφ1 , . . . , mφn−1 ) is a higher dimensional “power” and allows to look at
the dynamics of Tn,p (x) = p(x) and the corresponding Mandelbulb Mn,p . As with all celebrities,
there is a scandal:
63
Theorem: There is no theorem about the Mandelbulb Mn,m for n > 2.
Except of course the just stated theorem. But you decide whether it is true of not. The
Mandelbulb set has been discovered only recently. An attempt to trace some of its history
was done in [244]: already Rudy Rucker had experimented with a variant of M3,2 in 1988.
Jules Ruis wrote me to have written a Basic program in 1997. The first who wrote down
the formulas used today is Daniel White mentioned in a 2009 fractal forum. Jules Ruis 3D
printed the first models in 2010. See also [52] for some information on generating the graphics.
137. Banach Alaoglu

A Banach space X is a linear space equipped with a norm |·| defining a metric d(x, y) = |x−y|
with respect to which X is complete. The unit ball in X is the closed ball {x ∈ X | |x| ≤ 1}.
The dual space X ∗ of X is the linear space of linear functionals f : X → R with the norm
|f | = sup|x|≤1,x∈X |f (x)|. It is again a Banach space. The weak* topology is the smallest
topology on X ∗ which makes all maps f → f (x) continuous for all x ∈ X.
Theorem: The unit ball in a dual Banach space X ∗ is weak* compact.
The theorem was proven in 1932 in the separable case by Stefan Banach and in 1940 in general
by Leonidas Alaoglu. The result essentially follows from Tychonov’s theorem as X ∗ can be seen
as a closed subset of a product space. Banach-Alaoglu therefore relies on the axiom of choice. A
case which often appears in applications is when X = C(K) is the space of continuous functions
on a compact Hausdorff space K. In that case X ∗ is the space of signed measures on K. One
implication is that the set of probability measures is compact on K. An other example are
Lp spaces (p ∈ [1, ∞), for which the dual is Lq with 1/p + 1/q = 1 (meaning q = ∞ for p = 1)
and showing that for p = 2, the Hilbert space L2 is self-dual. In the work of Bourbaki the
theorem was extended from Banach spaces to locally convex spaces (linear spaces equipped
with a family of semi-norms). Examples are Fréchet spaces, (locally convex spaces which are
complete with respect to a translation-invariant metric). See [94].
138. Whitney trick

Let M be a smooth orientable simply connected d-manifold and two smooth connected sub-
manifolds K, L of dimension k and l such that k + l = d which have the property that K and L
intersect transversely in points x, y in the sense that the tangent spaces at the intersection
points span Tx M and Ty M and that they have opposite intersection sign. The two manifolds
K, L can be isotoped from each other along a disc if there exists a smooth 2-disk embedded
in M such that M ∩ K and M ∩ L are single points. The disk is called a Whitney disk. The
Whitney trick or Whitney lemma is:
Theorem: Any transverse K, L of ≥ 3 manifolds in M has a Whitney disk.
See [119]. See [270] for counter examples in d ≤ 4, who writes also “A hypothesis of algebraic
topology given by the signs of the intersection points leads to the existence of an isotopy”. The
failure of the Whitney trick in smaller dimensions is one reason why some questions in manifold
theory appear hardest in three or four dimension. There is a variant of the Whitney trick which
works also in dimensions 5, where K has dimension 2 and L has dimension 3.
64
OLIVER KNILL
139. Torsion groups

An elliptic curve E over Q is also called a rational elliptic curve. The curve E carries
an Abelian group structure where every addition of a point x → x + y is a morphism. The
torsion subgroup of E is the subgroup consisting of elements which all have finite order in
E. The Mordell-Weil theorem (which applies more generally for any Abelian variety) assures
that E = Zr ⊕ T , where T is a finite group and r is a finite number called the rank of E.
Mazur’s torsion theorem states that the only possible finite orders in E are 1, 2, 3, . . . , 9, 10
and 12. Only 15 different torsion subgroups appear in rational elliptic curves: Z1 , . . . , Z10 , Z12
or Z2 × Z2 , Z2 × Z4 , Z2 × Z6 and Z2 × Z8 . Lets call this collection of groups the Mazur class.
The theorem is:
Theorem: The torsion group of a rational elliptic curve is in the Mazur class.
The theorem was proven by Barry Mazur in 1977. [366].
140. Coloring
A graph G = (V, E) with vertex set V and edge set E is called planar if it can be embedded in
the Euclidean plane R2 without any of the edges intersecting. By a theorem of Kuratowski, this
is equivalent to a graph theoretical statement: G does not contain a homeomorphic image of
neither the complete graph K5 nor the bipartite utility graph K3,3 . A graph coloring with k
colors is a function f : V → {1, 2, . . . , k} with the property that if (x, y) ∈ E, then f (x) 6= f (y).
In other words, adjacent vertices must have different colors. The 4-color theorem is:
Theorem: Every planar graph can be colored with 4 colors.
Some graphs need 4 colors like a wheel graph having an odd number of spikes. There are
planar graphs which need less. The 1-point graph K1 needs only one color, trees needs only
2 colors and the graph K3 or any wheel graph with an even number of spikes only need 3
colors. The theorem has an interesting history: since August Ferdinand Möbius in 1840 spread
a precursor problem given to him by Benjamin Gotthold Weiske, the problem was first known
also as the Möbius-Weiske puzzle [376]. The actual problem was first posed in 1852 by
Francis Guthrie [287], after thinking about it with his brother Frederick, who communicated
it to his teacher Augustus de Morgan, a former teacher of Francis who told William Hamilton
about it. Arthur Cayley in 1878 put it first in print, (but it was still not in the language of
graph theory). Alfred Kempe published a proof in 1879. But a gap was noticed by Percy John
Heawood 11 years later in 1890. There were other unsuccessful attempts like one by Peter
Tait in 1880. After considerable theoretical work by various mathematicians including Charles
Pierce, George Birkhoff, Oswald Veblen, Philip Franklin, Hassler Whitney, Hugo Hadwiger,
Leonard Brooks, William Tutte, Yoshio Shimamoto, Heinrich Heesch, Karl Dürre or Walter
Stromquist, a computer assisted proof of the 4-color theorem was obtained by Ken Appel
and Wolfgang Haken in 1976. In 1997, Neil Robertson, Daniel Sanders, Paul Seymour, and
Robin Thomas wrote a new computer program. Goerge Gonthier produced in 2004 a fully
machine-checked proof of the four-color theorem [428]. There is a considerable literature like
[319, 43, 153, 356, 80, 428].
65
141. Contact Geometry

Assume M is a smooth compact orientable (2n − 1)-manifold equipped with an auxiliary Rie-
mannian metric g. A 1-form α ∈ Λ1 (M ) defines a field of hyperplanes ξ = ker(α) ⊂ T M .
Conversely, given a field of hyperplanes, one can define α = g(X, ·), where X is a local non-zero
section of the line bundle ξ ⊥ . A contact structure is a hyperplane field ξ = dα for which
the volume form α ∧ (dα)n is nowhere zero. The 1-form α is then called a contact form and
(M, ξ) is called a contact manifold. The Reeb vector field R is defined by dα(R, ·) = 0,
α(R) = 1. The Weinstein conjecture is a theorem in dimension 3:
Theorem: On a 3-manifold, the Reeb vector field has a closed periodic orbit.
The theorem was proven by Clifford Taubes in 2007 using Seiberg-Witten theory. Mike Hutch-
ings with Taubes established 2 Reeb orbits under the condition that all Reeb orbits R are
non-degenerate in the sense that the linearized flow does not have an eigenvalue 1. Hutch-
ings with Dan Cristofaro-Gardiner later removed the non-degeneracy
R condition [209, 103] and
also showed that if the product of the actions A(γ) = γ α of the two orbits is larger than the
R
volume M α ∧ dα of the contact form, then there are three. To the history: Alan Weinstein
has shown already that if Y is a convex compact hypersurface in R2n , then there is a periodic
orbit. Paul Rabinovitz extended it to star-shaped surfaces. Weinstein conjectured in 1978 that
every compact hypersurface of contact type in a symplectic manifold has a closed character-
istic. Contact geometry as an odd dimensional brother of symplectic geometry has become
its own field. Contact structures are the opposite of integrable hyperplane fields: the Frobe-
nius integrability condition α ∧ dα = 0 defines an integrable hyperplane field forming a
co-dimension 1 foliation of M . Contact geometry is therefore a “totally non-integrable hyper
plane field”. [157]. The higher dimensional case of the Weinstein conjecture is wide open [207].
Also the symplectic question whether every compact and regular energy surface H = c for a
Hamiltonian vector field in R2n has a periodic solution is open. One knows that there are for
almost all energy values in a small interval around c. [200].
142. Simplicial spheres

A convex polytope G is defined as the convex hull of n points in Rd such that all vertices
are extreme points called vertices. (Extreme points are points which do not lie in an open
line segment of G.) This definition of [176] is also called a polytopal sphere. A simplicial
sphere is a geometric realization of a simplicial complex that is homeomorphic to the standard
(d-1)-dimensional spheres in Rd . For a polytopal sphere, the boundary of G is made up of
(d − 1)-dimensional polytopes called (d − 1)-faces. A cyclic polytope C(n, d) can be realized
as the convex hull of the n vertices {(t, t2 , t3 , · · · td ) | t = 1, 2, . . . , n} ⊂ Rd . Let fk (G) denote
the number of k-dimensional faces in G. So, f0 (G) is the number of vertices, f1 (G) the number
of line segments and fd−1 the number of facets, the highest dimensional faces in G. Extending
the definition to f−1 = 1 (counting the empty complex, which is a (−1)-dimensional complex),
the vector f = (f−1 , f0 , f1 , · · · fd ) is called the extended f -vector of G. The upper bound
theorem is
Theorem: For simplicial spheres with f0 (G) = n, then fk (G) ≤ fk (C(n, d)).
This had been the upper bound conjecture of Theodore Motzkin from 1957 which was
proven by Peter McMullen in 1970 who reformulated it hk (G) ≤ n−d+k−1
k
for all k < d/2
66
OLIVER KNILL
as the other numbers are determined by Dehn-Sommerville conditions hk = hd−k for 0 ≤

k ≤ d. The h-vector (h0 , . . . hd ) and f -vector (f−1 , f0 , . . . , fd−1 ) determine each other via
Pd d−k
= dk=0 hk td−k . Victor Klee suggested the upper bound conjecture to be
P
k=0 fk−1 (t − 1)
true for simplicial spheres, which was then proven in by Richard Stanley in 1975 using new
ideas like relating hk with intersection cohomology of a projective toric variety associated
with the dual of G. (A toric variety is an algebraic variety containing an algebraic torus as
an open dense subset such that the group action on the torus extends to the variety.) The
result for simplicial spheres implies the result for convex polytopes because a subdivision of
faces of a convex polytope into simplices only increases the numbers fk . The g-conjecture of
McMullen from 1971 gives a complete characterization of f -vectors of simplicial spheres. Define
g0 = 1 and gk = hk − hk−1 for k ≤ d/2. The g-conjecture claims that (g0 , . . . g[d/2] ) appears as
a g-vector of a sphere triangulation if and only if there exists a multicomplex Γ with exactly
gk vectors of degree k for all 0 ≤ i ≤ [d/2]. (A multi-complex Γ is a set of non-negative
integer vectors (aP 1 , . . . , an ) such that if 0 ≤ bi ≤ ai , then (b1 , . . . bn ) is in Γ. The degree of a
multicomplex is i ai .) The g-theorem proves this for polytopal spheres (Billera and Lee in
1980 sufficiency) and (Stanley 1980 giving necessity). The g-conjecture is open for simplicial
spheres. [438, 381, 81]
143. Bertrand postulate

A basic result in number theory is
Theorem: For n > 1, there always exists a prime p between n and 2n.
As the theorem was conjectured in 1845 by Joseph Bertrand, it is still called Bertrand’s
postulate. Since Pafnuty Tschebyschef’s (Chebyshev) proof in 1852, it is a theorem. For
a proof, see [211] page 367. Srinivasa Ramanujan
P simplified Chebyshev’s proof considerably
in 1919 and strengthened it: if π(x) = p≤x,p prime 1 is the prime counting function, then
Bertrand’s result can be restated as π(x) − π(x/2) ≥ 1 for x ≥ 2. Ramanujna shows that
π(x) − π(x/2) ≥ k, for large enough x (larger or equal than pk ). The primes pk giving the
lower bound for x solving this are called Ramanujan primes. Simple proofs like one of Erdös
from 1932 are given in Wikipedia or [204] page 82, who notes ”it is not a very sharp result.
Deep analytic methods can be used to give much better results concerning the gaps between
successive primes”. There is a very simple proof assuming the Goldbach conjecture (stating
that every even number larger than 2 is a sum of two primes): [341] if n is not prime, then
2n = p + q is a sum of two primes, where one is larger than n and one smaller than 2n; on
the other hand, if n is prime, then n + 1 is not prime and 2n + 2 = p + q is a sum of two
primes, where one, say q is larger than n and smaller than 2n + 2. But q can not be 2n + 1 (as
that would mean p = 1), nor 2n (as 2n is composite) so that n < q < 2n. There are various
generalizations like Mohamed El Bachraoui’s 2006 theorem that there are primes between 2n
and 3n or Denis Hanson from 1973 [183] that there are primes between 3n and 4n for n ≥ 1.
Mohamed El Bachraoui asked in 2006 whether for all n > 1 and all k ≤ n, there exists a
prime in [kn, (k + 1)n] which is for k = 1 the Bertrand postulate. A positive answer would give
that there is always a prime in the interval [n2 , n2 + n]. Already the Legendre conjecture,
asking whether there is always a prime p satisfying n2 < p < (n + 1)2 for n ≥ 1 is open. The
Legendre’s conjecture is the fourth of the super famous great problems of Edmund Landau’s
1912 list: the other three are the Goldbach conjecture, the twin prime conjecture and
then the Landau conjecture asking whether there are infinitely many primes of the form
67
n2 + 1. Landau really nailed it. There are 4 conjectures only, but all of them can be stated
in half a dozen words, are completely elementary, and for more than 100 years, nobody has
proven nor disproved any of them.
144. Non-squeezing theorem

The Euclidean space M = R2n carries
the standard symplectic 2-form ω(v, w) = (v, Jw) with
0 I
the skew-symmetric matrix J = . A linear transformation f : M → M, x → Ax is
−I 0
called symplectic, if A satisfies AT JA = J. A smooth transformation f : M → M is called
a symplectomorphism if it is a diffeomorphism and if the derivative df is a symplectic map
from Tx M → Tf (x) M at every point x ∈ M . Any smooth map for which df is symplectic is
automatically a diffeomorphism as symplectic matrices have determinant 1 and are so invertible.
Let B(r) = {x ∈ M | x · x ≤ r2 } denote the round solid ball of radius r and Z(r) = {x ∈
M | x21 + y12 ≤ r2 } the solid cylinder of radius r. Given two sets A, B, one says there is a
symplectic embedding of A in B, if there exists a symplectomorphism f such that f (A) ⊂ B.
As symplectic maps are volume preserving, a necessary condition is Vol(A) ≤ Vol(B). Is this
the only constraint? Yes, for n = 1, where the cylinder and the ball are the same as defined
B(r) = Z(r). But no in higher dimensions n ≥ 2 by the Gromov non-squeezing theorem:
Theorem: A symplectic embedding B(r) → Z(R) implies r ≤ R.
The theorem has been proven in 1985 by Michael Gromov. It has been dubbed as the principle
of the symplectic camel by Maurice de Gosson referring to the “eye of the needle” metaphor.
A reformulation of the Gosson allegory [107] after encoding “camel” = “ball in the phase
space”, “hole = “cylinder”, and “pass”=“symplectically embed into”, “size of the hole” =
“radius of cylinder” and “size of the camel” = “radius of the ball” is: “There is no way that
a camel can pass through a hole if the size of the hole is smaller than the size of the camel”.
See [292, 208] for expositions. The non-squeezing theorem motivated also the introduction of
symplectic capacities, quantities which are monotone c(M ) ≤ c(N ) if there is a symplectic
embedding of M into N , which are conformal in the sense that if ω is scaled by λ, then c(M )
is scaled by |λ| and such that c(B(1)) = c(Z(1)) = π. For n = 1, the area is an example of a
symplectic capacity (actually unique). The existence of a symplectic capacity obviously proves
the squeezing theorem. Already Gromov introduced an example, the Gromov width, which
is the smallest. More are constructed in using calculus of variations. See [200, 293].
145. Kähler Geometry

A Kähler manifold is a complex manifold (M, J) together with a Hermitian metric h whose
associated Kähler form ω is closed. (The manifold can be given by a Riemannian metric g
compatible with the complex structure g(JX, JY ) = g(X, Y ). The Kähler form ω is then a
2-form ω(X, Y ) = g(JX, Y ) satisfying dω = 0 and h = g + iω is the Hermitian metric. (M, ω)
is then also a symplectic manifold.) As ω is closed, it represents an element in the cohomology
class H 2 (M ) called Kähler class. The Calabi inverse problem is: given a compact Kähler
manifold (M, ω0 ) and a (1, 1)-form R representing 2π times the first Chern class of M , find a
metric ω in the Kähler class of ω0 such that Ricci(ω) = R. In local coordinates, one can write
Ricci(ω) = −i∂∂ log det(g). For compact M :
68
OLIVER KNILL
Theorem: The Calabi inverse problem has a unique solution ω.

This was conjectured in 1957 by Eugenio Calabi and proven in 1978 by Shing-Tung Yau by
solving nonlinear Monge-Ampère equations using analytic techniques like Nash-Moser type
techniques. The theorem implies that if the first Chern class of M is zero, then (M, ω0 )
carries has a unique Ricci-flat Kähler metric g in the same Kähler class than ω0 . Kähler
geometry deals simultaneously with Riemannian, symplectic and complex structures: (M, g) is
a Riemannian, (M, ω) is a symplectic and (M, J) is a complex manifold. The inverse problem
of characterizing geometries from curvature data is central in all of differential geometry. Here
are some examples: a) M = C with Euclidean metric g is Kähler with ω = (i/2) k dz ∧ dz k
n k
P
but it is not compact. But if Γ is a lattice, then the induced metric on the torus Cn /Γ is
Kähler. b) Because complex submanifolds of a Kähler manifold are Kähler , and the complex
n
projective
P space2 CP with the Fubini-Study metric is Kähler (with ω = i∂∂ρ, where ρ =
log(1 + k |zk | /2) is the Kähler potential), any complex projective variety is Kähler.
d) For the complex hyperbolic case where M is the unit ball in Cn , the Kähler potential is
ρ = 1 − |z|2 . By Kodeira, Kähler forms representing an integral cohomology class correspond to
projective algebraic varieties. c) Calabi-Yau manifolds are complex Kähler manifolds with zero
first Chern classes. Examples are K3 surfaces. The existence theorem assures that they carry
a Ricci-flat metric, which are examples of Kähler-Einstein metrics. Also Hodge theory
works well for Kähler manifolds. In the complex, the Dolbeault operators ∂, ∂ and d = ∂ + ∂
lead to Hodge Laplacians ∆∂ , ∆∂ and ∆d , and so to harmonic forms H p,q P for differential
forms of type (p, q) and harmonic r-forms for ∆. In the Kähler case, H = p+q=r H p,q . An
r
example result due to Lichnerowicz is that if Ricci(Ω) ≥ λ > 0, then the first eigenvalue λ1 of
∆ satisfies λ1 ≥ 2λ. See [30, 424, 25].
146. Projective Geometry

A conic section is a curve which is obtained when intersecting a cone x2 +y 2 = z 2 with a plane
ax+by+cz = d. A bit more general is a conic, an algebraic curve az 2 +bxy+cy 2 +dx+ef +g = 0
of degree 2. They are either non-singular conics, classified as ellipses like x2 + y 2 = 1,
hyperbola x2 −y 2 = 1 or parabola x2 = y, or then degenerate conics like a point x2 +y 2 = 0,
the cross x2 = y 2 , the line x2 = 0 or pair of parallel lines x2 = 1. Given 6 different
points A1 , A2 , A3 , B1 , B2 , B3 on a conic, where A1 , A2 , A3 are neighboring and B1 , B2 , B3 are
neighboring, a Pascal configuration is the set of lines Ai Bj with i 6= j. The intersection
points of this Pascal configuration is the set of three intersections of Ai Bj with Aj Bi , where
{i, j} runs over all three 2-point subsets of {1, 2, 3}.
Theorem: The intersection points of a Pascal configuration are on a line.
The theorem was found in 1639 by Blaise Pascal (as a teenager) in the case of an ellipse. A
limiting case where we have two crossing lines is the Pappus hexagon theorem which goes
back to Pappus of Alexandria who lived around 320 AD. The Pappus hexagon theorem is one
of the first results in projective geometry.
147. Vitali theorem

A Lebesgue measure in Euclidean space Rn is a Borel measure which is invariant under
Euclidean transformations. It is the Haar measure of the locally compact group Rn and unique
if the unit cube has measure 1. In dimension n = 1, the Lebesgue measure of an interval [a, b]
69
is b − a. In dimension n = 2, the Lebesgue measure of a measurable set is the area of the set. In
particular, a ball of radius r has area πr2 . When constructing the measure one has to specify
a σ-algebra, which is in the Lebesgue case the Borel σ-algebra generated by the open sets in
Rn . One has for every n ≥ 1:
Theorem: There exist sets in Rn that are not Lebesgue measurable.
The result is due to Giusetppe Vitali from 1905. It justifies why one has to go through all the
trouble of building a σ-algebra carefully and why it is not possible to work with the σ-algebra
of all subsets of R (which is called the discrete σ-algebra). The proof of the Vitali theorem
shows connections with the foundations of mathematics: by the axiom of choice there exists
a set V which represents equivalence classes in T/Q, where T is the circle. For this Vitali set
V all translates Vr = V + r are all disjoint with r ∈ Q. {r + V, r ∈ Q} = R is a partition. By
the Lebesgue measure property, all translated sets Vr have the same measure. As they are a
countable set and are disjoint and add up to a set of Lebesgue measure 1, they have to have
measure zero. But this contradics σ additivity. Now lift V to R and then build V × Rn−1 .
More spectacular are decompositions of the unit ball into 5 disjoint sets which are equivalent
under Euclidean transformations and which can be reassembled to get two disjoint unit balls.
This is the Banach-Tarski construction from 1924.
148. Wilson’s theorem

The factorial n! of a number is defined as n! = 1 · 2 · · · n.
Theorem: n > 1 is prime if and only if (n − 1)! + 1 is divisible by n.
It is named after John Wilson, who was a student of Edward Waring. It seems that Joseph-Louis
Lagrange gave the first proof in 1771. It is not a practical way to determine primality: [384]:
from a computational point of view, it is probably one of the world’s least efficient primality
tests since computing (n − 1)! takes so many steps. Also named after him are Wilson primes,
primes for which not only p but p2 divides (p − 1)! + 1. The smallest one is 5. It is not known
whether there are infinitely many.
149. Carleson theorem

If f ∈ L2 (T), where T = R/(2πZ) P is the circle, then the Fourier transform L2 (T) → l2 (Z)
gives a Fourier series g(x) = k∈Z ck eikx , where c = (. . . , c−2 , c−1 , c0 , c1 , c2 , . . . ) ∈ l2 (Z) is
−1
R ikx
Rgiven2 by ck = (2π)
P 2 T f (x)e dx. For smooth f , one knows g = f and Parseval’s identity
R
f (x) dx = k ck so that the Fourier transform extends to an unitary operator L2 (T) →
l2 (Z). This does not say anything yet about the convergence of the sequence
Pn gn (x). We say the
Fourier series converges to f at a point x, if the sequence gn (x) = k=−n ck eikx converges to
f (x) for n → ∞. We say, a sequence gn (x) converges almost everywhere to f , if there exists a
set Y ⊂ T of full Lebesgue measure µ(T) = 1 such that the series converges for all x ∈ Y . (The
Lebesgue measure is the normalized Haar measure dx/(2π) on the circle). That the question
can be subtle is illustrated by the result of Andrey Kolmogorov from 1923 to 1926, who gave
examples of L1 (T) functions for which the Fourier series diverges everywhere.
Theorem: The Fourier series of a L2 function converges almost everywhere.
70
OLIVER KNILL
The statement had been conjectured by Nikolai Luzin in 1915 and was known as the Luzin
conjecture. The theorem was proven by Lennard Carleson in 1966. An extension to Lp with
p ∈ (1, ∞] was proven by Richard Hunt in 1968. The proof of the Carleson theorem is difficult.
While mentioned in harmonic analysis texts like [236] or surveys [238], who say about the
Carleson-Hunt theorem that it is one of the deepest and least understood parts of the theory.
150. Intermediate value

Let (X, O) be connected topological space and f : X → R a continuous map. We say f takes
both positive and negative signs if there exists a, b ∈ X such that f (a) < 0 nd f (b) > 0. A
root of f is a point x such that f (x) = 0. Let C(X) denote the set of continuous functions
from X to R, meaning that for f ∈ C(X) and all open sets U in R, one has f −1 (U ) ∈ O.
Theorem: f ∈ C(X) reaching both signs on a connected X has a root.
The theorem was proven by Bernard Bolzano in 1817 for functions from [a, b] to R. The proof
follows from the definitions: as P = (0, ∞) is open, also f −1 (P ) is open. As N = (−∞, 0)
is open, also f −1 (N ) is open. If there is no root, then X = N ∪ P is a disjoint union of
two open sets and so disconnected contradicting the assumption of X being connected. A
consequences is the wobbly table theorem: given a square table with parallel equal length
legs and a “floor” given by the graph z = g(x, y) of a continuous g can be rotated and possibly
translated in the z direction so that all 4 legs are on the table. The proof of this application
is seen as a consequence of the intermediate value theorem applied to the height function f (φ)
of the fourth leg if three other legs are on the floor. A consequence is also Rolle’s theorem
assuring that if a continuously differentiable function [a, b] → R with f (a) = f (b) has a point
x ∈ (a, b) with f 0 (x) = 0. Tilting Rolle gives the mean value theorem assuring that for a
continuously differentiable function [a, b] → R there exists x ∈ (a, b) with f 0 (x) = f (b) − f (a).
The general theorem shows that it is the connectedness and not completeness which is the
important assumption.
151. Perron-Frobenius
A n × n matrix A is non-negative if Aij ≥ 0 for all i, j and positive if Aij > 0 for all i, j.
The Perron-Frobenius theorem is:
Theorem: A positive matrix has a unique largest eigenvalue.
It has been proven by Oskar Perron in 1907 and by Georg Frobenius in 1912. When seeing
the map x → Ax on the projective space, this is in suitable coordinates a contraction and the
Banach fixed point theorem applies. The Brouwer fixed point theorem only gives existence, not
uniqueness, but the Brouwer fixed point applies for non-negative matrices. This has applications
in graph theory, Markov chains or Google page rank. The Google matrix is defined as G =
dA + (1 − d)E, where d is a damping factor and A is a Markov matrix defined by the network
and E is the matrix Eij = 1. Sergey Brin and Larry Page write “the damping factor d is the
probability at each page the random surfer will get bored and request another random page”.
The page rank equation is Gx = x. In other words, the Google Page rank vector (the one
billion dollar vector), is a Perron-Frobenius eigenvector. It assigns page rank values to the
individual nodes of the network. See [274]. For the linear algebra of non-negative matrices, see
[301].
71
152. Continuum hypothesis

ℵ0 is the cardinality of the natural numbers N. ℵ1 is the next larger cardinality. The cardi-
nality of the real numbers R is 2ℵ0 . The statement 2ℵ0 = ℵ1 is the continuum hypothesis
abbreviated CH. The Zermelo-Fraenkel axiom system ZFC of set theory is the most com-
mon foundational axiomatic framework of mathematics. The letter C refers to the axiom of
choice.
Theorem: 2ℵ0 = ℵ1 nor 2ℵ0 6= ℵ1 can be proven in ZFC.
This result is due to Paul Cohen from 1963. Cantor had for a long time tried to prove that
the continuum hypothesis holds. Cohen’s theorem shows that any such effort had been in vain
and why Cantor was doomed not to succeed. The problem had then been the first of Hilbert’s
problems of 1900. [364].
153. Homotopy-Homology
Given a path connected pointed topological space X with base b, the n’th homotopy group
πn (X) is the set of equivalence classes of base preserving maps from the pointed sphere S n to
X. It can be written as the set of homotopy classes of maps from the n-cube [0, 1]n to X
such that the boundary of [0, 1]n is mapped to b. It becomes a group by defining addition as
(f +g)(t1 , . . . , tn ) = f (2t1 , t2 , . . . tn ) for 0 ≤ t1 ≤ 1/2 and (f +g)(t1 , . . . , tn ) = g(2t1 −1, t2 , . . . , tn )
for 1/2 ≤ t ≤ 1. In the case n = 1, this is “joining the trip”: travel first along the first curve
with twice the speed, then take the second curve. The groups πn do not depend on the base
point. As X is assumed to be connected, π0 (X) is the trivial group. The group π1 (X) is
the fundamental group. It can be non-abelian. For n ≥ 2, the groups πn (X) are always
abelian f + g = g + f . The k’th homology group Hn (X) of a topological space X with
integer coefficients is obtained from the chain complex of the free abelian group generated by
continuous maps from n-dimensional simplices to X. The Hurewicz theorem is
Theorem: There exists a homomorphism πn (X) → Hn (X).
Higher homotopy groups were discovered by Witold Hurewitz from 1935-1936. The Hurewitz
theorem is due to Hurewicz from 1950 [206]. In the case n = 1 the homomorphism can be easily
described: if γ : [0, 1] → X is a path, then since [0, 1] is a 1-simplex, the path is a singular
1-simplex in X. As the boundary of γ is empty, this singular 1-simplex is a cycle. This allows
to see it as an element in H1 (X). If two paths are homotopic, then their corresponding singular
simplices are equivalent in H1 (X). There is an elegant proof using Hodge theory if X = M is a
compact manifold: the image C of a map πp (M ) can be interpreted as a Schwartz distribution
on M . Let L = (d + d∗ )2 be the Hodge Laplacian and let the heat flow e−tL act on C. For
t > 0, the image e−tL C is now smooth and defines a differential form in Λp (M ). As all the non-
zero eigenspaces get damped exponentially, the limit of the heat flow is a harmonic form,
an eigenvector to the eigenvalue 0. But Hodge theory identifies ker(L|Λp ) with H p (M ) and
so with Hp (M ) by Poincaré duality. The Hurewitz homomorphism is then even constructive.
“Just heat up the curve to get the corresponding cohomology element, the commutator group
elements get melted away by the heat.” A space X is called n-connected if πi (X) = 0 for all
i ≤ n. So, 0-connected means path connected and 1-connected is simply connected. For
n ≥ 2, one has πn (X) isomorphic to Hn (X) if X is (n − 1)-connected. In the case n = 1, this
can already not be true as π1 (X) is in general non-commutative and H1 (X) is but H1 (X) is
72
OLIVER KNILL
the isomorphic to the abelianization of G = π1 (X) which is the group obtained by factoring
out the commutator subgroup [G, G] which is a normal subgroup of G and generated by all the
commutators g −1 h−1 gh of group elements g, h of G. See [190].
154. Pick’s theorem

Let P be a simple polygon in the plane R2 . This means that it is given by as sequence of points
called vertices Pi = (xi , yi ) such that the line segments Pi Pi+1 called edges joining neighboring
points do not intersect. The polygon defines a polygonal region G with area A. Assume now
that all xi , yi are integers. Let I be the number of lattice points (k, l) ∈ Z2 inside G and B the
number of lattice points at the boundary of G. Pick’s theorem assures:
Theorem: A = I + B/2 − 1.
The result was found in 1899 by Georg Pick [329]. For a triangle for example with no interior
points, one has 0 + 3/2 − 1 = 1/2, for a rectangle parallel to the coordinate axes with I = n ∗ m
interior points and B = 2n + 2m + 4 boundary points and area A = (n + 1)(m + 1) also
I − B/2 − 1 = A. The theorem has become a popular school project assignment in early
geometry courses as there are many ways to prove it. An example is to cut away a triangle and
use induction on the area then verify that if two polygons are joined along a line segment, the
functional I +PB/2 − 1 is additive. There are other explicit formulas for the area like Green’s
formula A = n−1 i=0 xi yi+1 − xi+1 yi which does not assume the vertices Pi = (xi , yi ) to be lattice
points.
155. Isospectral drums

On a region G ⊂ R2 with piecewise smooth boundary δG one can look at the Dirichlet
problem −∆f = 0 in the interior of G and f = 0 on δG. The region is considered a “drum”.
If hit, one hears the spectrum of the Laplacian ∆u = uxx + uyy . There is a sequence of
eigenvalues 0 = λ0 < λ1 ≤ λ2 ≤ · · · , real values which solve −∆un = λn un for some functions
un which are zero on the boundary. For example, if G is the square [0, π] × [0, π], then the
eigenvalues are n2 + m2 with eigenvectors sin(nx) sin(mx). The eigenvalue 0 belongs to the
constant eigenfunction. Two drums are called isospectral, if they have the same eigenvalues.
Two drums are non-isometric, if there is no transformation generated by rotations, translation
and reflections which maps one drum to the other.
Theorem: There exist non-isometric but isospectral drums.
Marc Katz asked in 1962 “Can one hear the sound of a drum” [225]”. Caroline Gordon, David
Webb and Scott Wolpert answered negatively [166]. In the convex case, the question is still
open.
156. Bertrand theorem

The path r(t) of a particle in Rn moving in a central force potential V (x) = f (|x|) experi-
ences the central force F = −∇V (x) = −f 0 (|x|)x/|x|. In the case of the Newton potential
V (x) = −GM m/|x|, where the central mass M , the body mass m as well as the gravitational
constant G determines the force F (x) = −xGM m/|x|3 . The motion of the particle follows the
differential equations r00 (t) = −M Gr(t)/|r|3 , which conserve the energy E(r) = mr02 /2 + V (r)
and angular momentum L = mr ∧ r0 , a n(n + 1)/2 dimensional quantity. The invariance
73
of L assures that r(t) stays in the plane initially spanned by r(0) and r0 (0) and that the area
of the parallelogram spanned by r(t) and r0 (t) is constant. To see the natural potential in Rn
is, one has to go beyond Newton and pass to Gauss, who wrote the gravitational law in the
form div(F ) = 4πρ, where ρ is the mass density. It expresses that mass is the source for the
force field F . To get the force field in a central symmetric mass distribution, one can use the
divergence theorem in Rn and relate the integral of 4πρ over a ball of radius r with the flux
of F through the sphere S(r) of radius r. The former is 4πM , where M is the total mass in
the ball, the later is −|S(r)|F (r), where |S(r)| is the surface area of the sphere and the nega-
tive sign is because for an attractive force F (r) points inside. So, in three dimensions, Gauss
recovers the Newton gravitational law F (r) = −4πGM/|S(r)| = −GM/|r|2 . There is a natural
central force Kepler problem in any dimensions: in Rn , we have F (r) = −Cn r/|r|n where Cn
is a constant. For n = 1, there is a constant force pulling the particle towards the center, for
n = 2, one has a 1/|r| force which corresponds to a logarithmic potential, for n = 3, it is the
Newtonian inverse square 1/r2 force, in n = 4, it is a 1/r3 force. For n = 0, one formally gets
the harmonic oscillator which is Hook’s law. Which potentials lead to periodic motion?
The answer is surprising and was given by Bertrand: only the harmonic oscillator potential and
the Newtonian potential in R3 work. Let us call a central force potential all periodic if every
bounded (position and velocity) solution r(t) of the differential equations is periodic. Already
for the Kepler problem, there are not only motions on ellipses but also scattering solutions
moving on parabola or hyperbola, or then suicide motions, with r0 (0) = 0, where the particle
dives into the singularity.
Theorem: Only the Newton potentials for n = −1 and n = 3 are all periodic.
This theorem of Joseph Bertrand from 1873 tells that three dimensional space is special as it
in any other dimension, calendars would be almost periodic. We could live with that but there
are more compelling reasons why n = 3 is dynamically better: in other dimensions, only very
special orbits stay bounded. A small perturbation leads to the planet colliding with the sun
or escaping to infinity, both not very pleasant for possible inhabitants. Gauss’s analysis also
shows in any dimension n, the motion of a particle in a constant mass density is always given
by the harmonic oscillator motion, independent of the dimension: the divergence theorem
gives 4πρ|B(r)| = −|S(r)|F (r), where |B(r)| is the volume of the solid ball of radius r. This
gives the Hook law force F (x) = −4πρx/(n), where n is the dimension.
157. Catastrophe theory

Catastrophe theory describes the singularity structure of smooth functions f on a n-manifold M
parametrized by some r parameters. A basic assumption is that configurations of interest
of the functional f are critical points of f : M → R. Especially interesting are minima, stable
configurations. When changing parameters of f , bifurcations, structural changes of the critical
set can happen. Especially, minima can change their nature or disappear. In particular, the
function ft (xt )), where xt is a local minimum can change discontinuously, even if the function
(t, x) → ft (x) is smooth. Such discontinuous changes are called catastrophes. The stage for
Thom’s theorem is a smooth function f : Rn → Rr . One can think of f as a r parameter family
of functions on space Rn . Let ∇x = (∂x1 , . . . ∂xn ) is the gradient operator with respect to
the space variables and Mf = {(x, y) ∈ Rn × Rr | ∇x f = 0} is the submanifold on which points
are critical. The space X = C ∞ (Rn × Rr ) of smooth functions in space and parameter can be
equipped with the Whitney topology, the topology generated by a basis which is the union
74
OLIVER KNILL
of all the basis sets of C k Whitney topologies. A basis for the later is the set of all functions for
which f (j) (x, y) ∈ Uj for all 0 ≤ j ≤ k and U0 , · · · Uk are all open intervals. With the Whitney
C ∞ topology, X is a Baire space so that residual sets (countable intersections of open dense
sets), are dense. The next theorem works n = 2, r ≤ 6 and for n ≥ 3 if r ≤ 5 [299]
Theorem: For a residual set in X, Mf is an r-dimensional manifold.
The theorem was due to René Thom who initiated Catastrophe theory in a 1966 article and
wrote [404] building on previous work by Hassler Whitney. More work and proofs were done
by various mathematicians like John Mather or Bernard Malgrange. There is more to it: the
restriction Xf of the projection of the singularity set Mf onto the parameter space Rr can be
classified. Thom proved that that for r = 4, there are exactly seven elementary catastro-
phes: ‘fold”, “cusp”, “swallowtail”, “butterfly”, “hyperbolic umbillic”, “elliptic umbillic” and
“parabolic umbillic”. For r = 5, the number of catastrophe types is 11. The subject is part of
singularity theory of differentiable maps, a theory that started by Hassler Whitney in 1955.
The theory of bifurcations was developed by Henri Poincare and Alexander Andronov. See
also [299, 335, 410]. It is also widely studied in the context of dynamical systems [303].
158. Phase transition

Given a finite simple graph G = (V, E), an interactionP function J : PE → R and a scalar
field h : V → R defines a Hamiltonian H(σ) = (i,j)∈E Jij σi σj − µ j∈V hj σj on the set
of all functions σ : V → {−1, 1}. The interpretation is that σi are spin values, hj an
external magnetic field and Jij is an interaction function. The additional parameter
µ is a magnetic moment. The energy H defines a probability measure P on the set Ω =
{−1, 1}V of all spin configurations. It is the Gibbs-Bolzmann distribution P [{σ}] = e−H(σ) /Z,
where Z is is the normalization constant rendering P a probability measure. One calls Z the
partition function (as it is usually considered to be a function of some of the parameters
like temperature). Given a random variable =observable X : Ω → R, one is interested in the
expectation E[X]. An example is X(σ) = σi σj , which leads to the correlation. When replacing
H with βH, where β = 1/(KT ) is an inverse temperature parameter (T is the temperature
and K the Bolzmann constant), one can study the expectation of a random variable X in
dependence of β. One writes now also E[X] = hXiβ to stress the dependence on β. In the case
d
when
P G is a d-dimensional lattice G = [−L, L] , where two lattice points x, y are connected
if k |xk − yk | = 1 one look at the L → ∞ limit where G = Zd . In the case J = 1, h = 0
this is the Ising model. As J is positive, this is a Ferromagnetic situation. A parameter
value, where a quantity like Zβ or a derivative of it changes discontinuously is called a phase
transition.
Theorem: The Ising model in two dimensions has a phase transition.
This was first proven by Lars Onsager in 1944, who in a tour de force gave analytical solutions.
The analysis shows that there is a phase transition. The temperature T at which this happens
is called the Curie temperature. The one dimensional case had been solved by Ernst Ising
in 1925, who got it as a PhD project from his adviser Wilhelm Lenz. In one dimensions, there
is no phase transition. In three and higher dimensions, there are no analytical solutions. The
Ising model is only one of many models and generalizations. If the Jij are random one deals
with disordered systems. An example is the Edwards-Anderson model, where Jij are
Gaussian random variables. This is an example of a spin glass model. An other example is
75
the Sherrington-Kirkpatrick model from 1975, where the lattice is replaced by a complete
graph and the Jij define a random matrix. An other possibility is to change the spin to Zn
or the symmetric group (Potts) or then some other Lie group (Lattice gauge fields) and then
use a character to get a numerical value. Or one replaces the zero dimensional sphere with a
higher dimensional sphere like S 2 and takes σi · σj (Heisenberg model). See [368].
159. Ceva theorem

Given a triangle ABC in the Euclidean plane and a point O in the interior. For any choice of
points A0 on the segment BC, point B 0 on the segment AC and point C 0 on the segment AB,
one can look at the ratios r(AB) = AC 0 /C 0 B and r(BC) = BA0 /A0 C and r(CA) = CB 0 /B 0 A
in which the points bisect the sides of the triangle.
Theorem: r(AB)r(BC)r(CA) = 1
The theorem is called after Giovanni Ceva who wrote it down in 1678. Al-Mu’taman ibn Hud
from Zaragoza proved it already in the 11’th century. [202]. See [353].
160. Angle theorem

Given a circle C with center M and two points A, B on C. If P is a point on C, then AP B is
constant for all P in C which are on the same side than M with respect to the segment AB.
The angle AP B is called the inscribed angle of the secant AB. The next theorem is also
called the inscribed angle theorem.
Theorem: The angle AP B is half the angle AM B.
The theorem is believed to be known already to Thales of Miletus who is the first Greek
mathematician known by name (624 - 546 BC). It it is usually called Thales theorem in the
special case is if A, B are on a diagonal. Then the angle AP B is a right angle. A consequence
of the theorem is that the opposite angles of a quadrilateral which is inscribed in a circle add
up to π. Unlike the special case of the right angle which immediately follows from symmetry,
the full version of Thales theorem can surprise at first.
161. Total curvature

A smooth simple closed curve C in R3 is called a knot. If r(t) is the parametrization, then
0 00 0 3
R 2π = |r (t) × r (t)|/|r (t)| is the curvature of the parametrization of r. The integral K(C) =
κ(t)
0
κ(t) dt is the total curvature of r. We say C is unknotted if C can be deformed to a
circle S = {x2 + y 2 = 1, z = 0} = {r1 (t) = (cos(t), sin(t), 0), t ∈ [0, 2π]} meaning that there
exists a smooth function R(t, s) such that R(t, 0) = r(t) and R(t, 1) = r1 (t) such that for any
s, the curve Ct : t → R(t, s) is a simple closed curve.
Theorem: If C is a knot and K(C) ≤ 4π, then K is unknotted.
This is the theorem of Fary-Milnor, proven by Fary in 1949 and Milnor in 1950. The theorem
follows also from the existence of quadrisecants, which are lines intersecting the knot in 4
points [110]. The existence of quadrisecants was proven by Erika Pannwitz in 1933 for smooth
knots and generalized in 1997 by Greg Kuperberg to tame knots, knots which are equivalent
to polygonal knots.
76
OLIVER KNILL
162. Morley’s Theorem

An angle trisector of an angle α = ∠(CAB) in R2 is a pair of lines P A, QA through A such
that the angles ∠(CAP ), ∠(P AQ), ∠(QAB) are all equal. Given a triangle ABC, we can look
at the angle trisectors at each point and intersect the adjacent trisectors, leading to a triangle
P QR inside the triangle. The triangle P QR is called the Morley triangle of ABC. Morley’s
theorem is
Theorem: For any triangle ABC, the Morley triangle is equilateral.
Morley’s theorem was discovered in 1899 by Frank Morley. A short proof was given in 1995
by John H. Conway: assume the triangle ABC had angles 3α, 3β, 3γ so that α + β + γ = π/3.
Start with an equilateral triangle P QR of length 1. Build three triangles P QA with angles
β + π/3, α, γ + π/3, QCA with angles α + π/3, γ, β + π/3 and a triangle RBQ with angles
γ+π/3, β, α+π/3. Then fill in three other triangles ACQ, CBR, BAP with angles α, γ, β+2π/3
and γ, β, α + 2π/3 β, α, γ + 2π/3. These triangles fits together to a triangle of the shape ABC.
See [136].
163. Rising sun lemma

For g ∈ C([a, b]), we say the set E(g) = {x ∈ (a, b) | g(t) > g(x) for t > x} has the rising sun
property if E isS open, empty if and only if g is decreasing and if not empty, then E can be
written as E = n (an , bn ) with pairwise disjoint intervals with g(an ) ≤ g(bn ). See [39].
Theorem: f ∈ C([a, b], R) has the rising sun property.
The theorem is due to F. Riesz. The name “rising sun lemma” appeared according to [39] first
in [21]. The picture is to draw the graph of the function f . If light comes from a distant source
parallel to the x-axis, then the intervals (an , bn ) delimit the hollows that remain in the shade
at the moment of sunrise. The lemma is used to prove that every monotone non-decreasing
function is almost everywhere differentiable.
164. Uniform continuity

Uniform continuity is a stronger version of continuity. But unlike continuity, which is defined
for maps between topological spaces, uniform continuity needs more structure like metric spaces
or more generally topological spaces with a uniform structure. Given two metric spaces X
and Y , a function f : X → Y is called continuous if f −1 (U ) is open for every open U in Y .
A function f is called uniformly continuous if there exists a sequence of numbers Mn → 0
such that for every n, the condition d(x, y) ≤ 1/n implies that d(f (x), f (y)) ≤ Mn .
Theorem: For compact X, continuous implies uniformly continuous.

The theorem is due to Eduard Heine and Georg Cantor. Heine is known also for the Heine-Borel
theorem which states that in Euclidean spaces, the class of closed and bounded sets agrees with
the class of compact sets. The proof of the Heine-Cantor theorem uses the extreme value
theorem assuring that a continuous function on a compact space X achieves a maximum. Look
for every n and every x at the minimal Mn (x) such that if |x − y| ≤ 1/n, then |f (x) − f (y)| ≤
Mn (x). Now Mn (x) is non-negative and finite and depends continuously on x. By the extremal
value theorem there is a maximum. We call it Mn . This assures now that if |x − y| ≤ 1/n, then
77
|f (x) − f (y)| ≤ Mn . The Bolzano-Weierstrass or sequential compactness theorem assures

that a bounded sequence in Rn has a convergent subsequence. This is used in the intermediate
value theorem assuring that if f (a) < 0 and f (b) > 0, then there is an x with f (x) = 0. The
Heine-Cantor theorem together with the intermediate value theorem assures that continuous
functions are Riemann integrable. The additional uniform structure or metric structure is
also necessary when defining completeness in the sense that every Cauchy sequence converges.
Completeness is not a property of topological spaces: (0, 1) is not complete but R is complete
even so the two spaces are homeomorphic.
165. Jordan normal form

A n×n matrix A is similar to an other n×n matrix B if there exists an invertible n×n matrix
S such that B = S −1 AS. A matrix is in Jordan normal form or Jordan canonical form
if it is block diagonal, where each block is a Jordan block. A m × m matrix J is a Jordan
block, if Je1 = λe1 , and Jek =λek + ek+1 for k = 2, . . . , m. An example of a 3 × 3 Jordan
3 1 0
block matrix is J = 0 3 1 . In other words, A is of the form A = λ1 + N , where N is

0 0 3
m
nilpotent: N = 0 and more precisely only has 1 in the super diagonal above the diagonal.
Theorem: Every n × n matrix is similar to a matrix in Jordan normal form.
Up to the order of the Jordan blocks, the Jordan normal form is unique. If each Jordan block
is a 1 × 1 matrix, then the matrix is called diagonalizable. The spectral theorem assures
that a normal matrix AA∗ = ∗
A A is diagonalizable. Not every matrix is diagonalizable as the
1 1
shear matrix A = , a 2 × 2 Jordan block, shows. The theorem has been stated first
0 1
by Camille Jordan in 1870. For history, see [55]. The Jordan-Chevalley generalization states
that over an arbitrary perfect field, a matrix is similar to B + N , where B is semi-simple and
N is nilpotent and BN = N B. (See [205] page 17). A matrix B is called semi-simple if every
B-invariant linear subspace V has a complementary B-invariant subspace. For algebraically
closed fields, semi-simple is equivalent to be conjugated to a diagonal matrix. To the condition
on the field: a field k is called perfect if every irreducible polynomial over k has distinct roots.
166. Hippocrates theorem

The theorem is also called the lunes of Hippocrates or the lunes of Alhazen is a theorem
in planar geometry. Given a triangle ABC with right angle β, one can draw the circles with
diameter AC, AB and BC. They define two moon shaped regions U, V bounded by circles.
Theorem: The area of U plus the area of V is the area of the triangle.
The proof directly follows from Pythagoras just by relating the areas of half discs and triangle.
The result is remarkable as it was historically the first attempt for the quadrature of the
circle. The lunes are bound by circles, while the triangle is bound by line segments. The
theorem does the quadrature of the lunes. Hippocrates of Chios lived from about 470 to
410 BC. For history see [29] page 37.
78
OLIVER KNILL
167. Fermat-Hamilton principle

A point x is called a critical point of a differentiable function f : Rm → R, if ∇f (x) = 0,
where ∇f is the gradient of f . A point x0 is called a local maximum of f if there exists
r > 0 such that f (x) ≤ f (x0 ) for all |x − x0 | < r. The local maximum does not have to be
isolated. For a constant function, every point is a local maximum. The local maximum also
does not have to be a global maximum. The function f (x) = x4 − x2 has a local maximum
at x = 0 but this is not a global maximum because f (2) > f (0).
Theorem: If x0 is a local maximum of f , then ∇f (x0 ) = 0.

This generalizes to theR calculus of variations, where ∇f is replaced by the variation. In
b
the case when f (x) = a L(x(t), x0 (t)) dt is a function on the space of curves [a, b] → Rn (one
calls this then a functional or action functional) then we an look at the problem to minimize
the action. In that case, the gradient is δS = Lx (x(t), x0 (t)) − dtd Lx0 (x(t), x0 (t)) = 0. This so
called Hamilton principle can be seen as a generalization of the Fermat principle to infinite
dimensions. The equations δS = 0 are called the Euler-Lagrange equations or Lagrange
equations of the second kind. They are the starting point of Lagrangian mechanics.
Fermat’s original paper deals with the single variable situation but the higher dimensional
situation is similar. Fermat in some sense already looked at the action principle which is the
situation to minimize the arc length of a path in a medium with two different properties like
water and air. In that case the shortest path is described by the Fermat law or Fermat’s
principle.
168. Alternating sign

An alternating sign matrix is a square matrix with entries 0, 1, −1 such that the sum of
each row and column is 1 and the nonzero entries in each row and column alternate in sign.
Theorem: The number of n × n alternating sign matrices is n−1 (3k+1)!

Q
k=0 (n+k)! .
The numbers n−1

Q
k=0 (3k +1)!/(n+k)! are known as the Robbins numbers or Andrews-Mills-
Robbins-Rumsey numbers and are the integer sequence A005130 [1]. The alternating sign
conjecture was popularized by David Robbins in [344]. The theorem was proven by Doron
Zeilberger in 1994 [437]. A short proof was given by Greg Kuperberg in 1996 [269]. A book
about it is [60].
169. Combinatorial convexity

A finite set P of points in Rd is called r-convex, if there is a partition of P into r sets such
that their convex hulls intersect simultaneously in a non-empty set. Tverberg’s theorem states:
Theorem: A set of (r − 1)(d + 1) + 1 points in Rd is r-convex.

The decomposition of P into r subsets is called the Tverberg partition. In the one-dimensional
case d = 1, the theorem assures that 2r − 1 points on the line are r-convex. For r = 3 for exam-
ple, this means that 5 points are 3-convex. If the points are arranged x1 < x2 < x3 < x4 < x5 ,
the Tverberg partition {x1 , x4 }, {x2 , x5 }, {x3 }}. For r = 2, it implies Radon’s theorem which
tells that d + 2 points in Rd can be partitioned into 2 sets whose convex hulls intersect. For
example, 4 points {x1 , x2 , x3 , x4 } in R2 can be partitioned into two sets such that their convex
79
hull intersect. Indeed, the 4 points define a quadrilateral and the partition {{x1 , x3 }, {x2 , x4 }}
define the two diagonals of the quadrilateral.
The theorem has been proven by Helge Tverberg in 1966. See [407, 210].
170. The Umlaufsatz

Let r be a continuously differentiable curve in R2 . If r(t) is a parametrization for which the
speed is 1, we have r0 (t) = (cos(α(t)), sin(α(t)))
R 2π and a signed curvature κ(t) = α0 (t). If
[0, 2π] is the parameter interval, then K = 0 κ(t) dt is the total curvature change. The Hopf
Umlaufsatz is:
Theorem: For r ∈ C 1 , the total curvature is 2π.
The paper was proven in 1935 by Heinz Hopf [203] in a marvelous homotopy proof: define
f (s, t) as the argument of the line through r(s) and r(t) or continuously extend it s = t as the
argument of the tangent line. The direct line from (0, 0) to (1, 1) in the parameter st-plane
gives a total angle change of n2π where n is an integer. Now deform the curve from (0, 0) to
(1, 1) so that it first goes straight from (0, 0) to (0, 1), then straight from (0, 1) to (1, 1). Both
lines produce a deformation of π and show that n = 1. The theorem can be generalized to a
Gauss-Bonnet theorem for planar regions G. The total curvature of the boundary is 2π times
the Euler characteristic of G. For a discrete version, see [249].
171. Frobenius determinant

The Frobenius determinant theorem tells how the determinant of the “multiplication table
matrix” factors into irreducible polynomials: if G = {g1 , . . . , gn } is a finite group and xi is a
variable associated to the group element gi , then the matrix Aij == xgi gj satisfies
Qr
Theorem: det(A) = j=1 pj (x1 , . . . , xn ))dj
Here dj = deg(pj ) and r is the number of conjugacy classes of G. For an Abelian group, there
are n conjugacy classes. The theorem had been conjectured in 1896 by Richard Dedekind.
Frobenius proved it. [192, 91]
172. Koenig’s theorem

A matching M in a finite simple graph G = (V, E) is a subset M of the edges ES in which no two
edges have a common vertex. A vertex cover C is a set of vertices such that x∈C S(x) = V ,
where S(x) is the unit sphere of a vertex x. A bipartite graph is a graph for which V = V1 ∪V2
can be partitioned into two disjoint sets V1 , V2 such that all edges connect vertices from different
sets. König’s theorem, from 1931, also called König-Egeváry theorem is:
Theorem: If G is bipartite the maximal matching number is the minimal
vertex cover number.
The vertex cover problem, the problem to find a minimal vertex cover is a classical NP
complete problem. For example, for a cyclic graph G = C10 with 2n vertices {1, 2, 3, · · · , 2n}
which is bipartite, the set C = {2, 4, · · · 2n} is a minimal vertex cover. The edges M =
{(1, 2), (3, 4), · · · (2n − 1, 2n)} are a minimal matching. The example of an odd cyclic graph like
C9 already shows the bipartite condition is necessary. For C9 , the set {1, 3, 5, 7, 8} is a minimal
80
OLIVER KNILL
cover and M = {(1, 2), (3, 4), (5, 6), (7, 8)} is a minimal matching.
The origin of the theorem is attributed to Dénes Kn̈ig, who proved it in 1931 and wrote a
precursor paper in 1916 where he prove that a regular (constant vertex degree) bipartite graph
has a perfect matching (a matching which covers all vertices). For a proof see for example [113]
(Chapter 2).
Epilogue: Value
Which mathematical theorems are the most important ones? This is a complicated variational
problem because it is a general and fundamental problem in economics to define “value”. The
difficulty with the concept is that “value” is often a matter of taste or fashion or social influence
and so an equilibrium of a complex social system. Value can change rapidly, sometimes
triggered by small things. The reason is that the notion of value like in game theory depends
on how it is valued by others. A fundamental principle of catastrophe theory is that maxima
of a functional can depend discontinuously on parameter. As value is often a social concept,
this can be especially brutal or lead to unexpected viral effects. First of all, value is often
linked to historical or morale considerations. We tend more and more to link artistic and
scientific value also to the person. In mathematics, the work of Oswald Teichmüller or Ludwig
Bieberbach for example are linked to their political view and so devalued despite their brilliance
[360]. This happens also outside of science, in art or industry. The value of a company now
also depends on what “investors think” or what analysts see for potential gain in the future.
Social media try to measure value using “likes” or “number of followers”. A majority vote is a
measure but how well can it predict correctly what be valuable in the future? Majority votes
taken over longer times would give a more reliable value functional. Assume one could persuade
every mathematician to give a list of the two dozen most fundamental theorems and do that
every couple of years, and reflect the “wisdom of an educated crowd”, one could probably get a
pretty good value functional. Ranking theorems and results in mathematics are a mathematical
optimization problem by itself. One could use techniques known in the “search industry”. One
idea is to look at the finite graph in which the theorems are the nodes and where two theorems
are related with each other if one can be deduced from the other (or alternatively connect them
if one influences the other strongly). One can then run a page rank algorithm [274] to see
which ones are important. Running this in each of the major mathematical fields could give
an algorithm to determine which theorems deserve the name to be “fundamental”.
Opinions
Having taught a course called “Math from a historical perspective” at the Harvard extension
school motivated to write up the present document. This course Math E 320 is a rather
pedestrian but pretty comprehensive stroll through all of mathematics. At the end of the
course, students were asked as part of a project to write some short stories about theorems
or mathematical fields or mathematical persons. The present document benefits already from
these writings and also serves a bit as preparation for the course. It is interesting to see what
others consider important. Sometimes, seeing what others feel can change your own view. I
was definitely influenced by students, teachers, colleagues and literature as well of course by
the limitations of my own understanding. And my point of view has already changed while
writing the actual theorems down. Value is more like an equilibrium of many different factors.
81
In mathematics, values have changed rapidly over time. And mathematics can describe the
rate of change of value [335]. Major changes in appreciation for mathematical topics came
definitely at the time of Euclid, then at the time when calculus was developed by Newton and
Leibniz. Also the development of more abstract algebraic constructs or topological notions,
the start of set theory changed things considerably. In more modern time, the categorization
of mathematics and the development of rather general and abstract new objects, (for example
with new approaches taken by Grothendieck) changed the landscape. In most of the new
development, I remain the puzzled tourist wondering how large the world of mathematics
is. It has become so large that continents have emerged: we have applied mathematics,
mathematical physics, statistics, computer science and economics which have drifted
away to independent subjects and departments. Classical mathematicians like Euler would
now be called applied mathematicians, de Moivre would maybe be stamped as a statistician,
Newton a mathematical physicist and Turing a computer scientist and von Neuman an
economist.
Search
A couple of months ago, when looking for “George Green”, the first hit in a search engine
would be a 22 year old soccer player. (This was not a search bubble thing [325] as it was
tested with cleared browser cache and via anonymous VPN from other locations, where the
search engine can not determine the identity of the user). Now, I love soccer, played it myself
a lot as a kid and also like to watch it on screen, but is the English soccer player George
William Athelston Green really more “relevant” than the British mathematician George Green,
who made fundamental break through discoveries which are used in mathematics and physics?
Shortly after I had tweeted about this strange ranking on December 27, 2017, the page rank
algorithm must have been adapted, because already on January 4th, 2018, the Mathematician
George Green appeared first (again not a search bubble phenomenon, where the search engine
adapts to the users taste and adjusts the search to their preferences). It is not impossible that
my tweet has reached, meandering through social media, some search engine engineer who was
able to rectify the injustice done to the miller and mathematician George Green. The theory
of networks shows “small world phenomena” [418, 32, 417] can explain that such influences or
synchronizations are not that impossible [394]. But coincidences can also be deceiving. Humans
just tend to observe coincidences even so there might be a perfectly mathematical explanation
prototyped by the birthday paradox. See [291]. But one must also understand that search
needs to serve the majority. For a general public, a particular subject like mathematics is
not that important. When searching for “Hardy” for example, it is not Godfrey Hardy who
is mentioned first as a person belonging to that keyword but Tom Hardy, an English actor.
This obviously serves most of the searches better. As this might infuriate particular groups
(here mathematicians), search engines have started to adapt the searches to the user, giving the
search some context which is an important ingredient in artificial intelligence. The problem
is the search bubble phenomenon which runs hard against objectivity. Textbooks of the future
might adapt their language, difficulty and even their citations or the historical credit on who
reads it. Novels might adapt the language to the age of the user, the country where the user
lives, and the ending might depend on personal preferences or even the medical history of the
user (the medical history of course being accessible by the book seller via ‘big data” analysis
of user behavior and tracking which is not SciFi this is already happening). Many computer
games are already customizable as such. A person flagged as sensitive or a young child might
82
OLIVER KNILL
be served a ending rather than ending the novel in an ambivalent limbo or even a disaster.
[325] explains the difficulty. The issues have amplified even more since that book came out and
even influenced elections.
Beauty
In order to determine what is a “fundamental theorem”, also aesthetic values matter. But the
question of “what is beautiful” is even trickier. Many have tried to define and investigate the
mechanisms of beauty: [185, 422, 423, 348, 373, 7, 304]. In the context of mathematical for-
mulas, the question has been investigated within the field of neuro-aesthetics. Psychologists,
in collaboration with mathematicians have measured the brain activity of 16 mathematicians
with the goal to determine what they consider beautiful [355]. The Euler identity eiπ + 1 = 0
was rated high with a value 0.8667 while a formula for 1/π due to Ramanujan was rated low
with an average rating of -9.7333. Obviously, what mattered was not only the complexity of
the formula but also how much insight the participants got when looking at the equation. The
authors of that paper cite Plato who wrote once ”nothing without understanding would ever
be more beauteous than with understanding”. Obviously, the formula of Ramanujan is much
deeper but it requires some background knowledge for being appreciated. But the authors
acknowledge in the discussion that that correlating “beauty and understanding” can be tricky.
Rota [348] notes that the appreciation of mathematical beauty in some statement requires the
ability to understand it. And [304] notices that “even professional mathematicians specialized
in a certain field might find results or proofs in other fields obscure” but that this is not much
different from say music, where “knowledge bout technical details such as the differences be-
tween things like cadences, progressions or chords changes the way we appreciate music” and
that “the symmetry of a fugue or a sonata are simply invisible without a certain technical
knowledge”. As history has shown, there were also always “artistic connections” [155, 66] as
well as “religious influences” [281, 374]. The book [155] cites Einstein who defines “mathematics
as the poetry of logical ideas”. It also provides many examples and illustrations and quotations.
And there are various opinions like Rota who argues that beauty is a rather objective property
which depends on historic-social contexts. And then there is taste: what is more appealing, the
element of surprise like the Birthday paradox or Petersburg paradox in probability theory, the
Banach-Tarski paradox in measure theory which obviously does not trigger any enlight-
enment nor understanding if one hears the first time: one can disassemble a sphere into 5
pieces, rotate and translate these pieces in space to build up two spheres. Or the surprising fact
that the infinite sum 1 + 2 + 3 + 4 + 5 + . . . is naturally equal to −1/12 as it is ζ(−1) (which
is defined by analytic continuation). The role of aesthetic in mathematics is especially impor-
tant in education, where mathematical models [148], mathematical visualization [31], artistic
enrichment [149], surfaces [266], or 3D printing [361, 256] can help to make mathematics more
approachable.
The fate of fame

Aesthetics is a fragile subject however. If something beautiful has become too popular and so
entered pop-culture, a natural aversion against it can develop. It is in danger to become a
clishé or even become kitsch (which is a word used to tear down popular stuff or label poor
taste). The Mandelbrot set for example is just marvelous, but it does not excite anymore
because it is so commonly known. The Monty-Hall problem which became famous by
Gardner columns in the nineties (see [375, 347]) was cool to teach in 1994, three years after
83
the infamous “parade column” of 1991 by Marilyn vos Savant. Especially after a cameo in the
movie “21”, the theorem has become part of mathematical kitsch. I myself love mathematical
kitsch. A topic that gained that status must have been nice and innovative to obtain that label.
Kitsch becomes only tiresome however if it is not presented in a new and original form. The book
[326] for example, in the context of complex dynamics, remains a master piece still today, even-
so the picture have become only too familiar. Rendering the mandelbrot set today however
does not rock the boat any more. In that context, it appears strange that mathematicians
do not jump on the “mandelbulb set” M which is one of the most beautiful mathematical
objects there is but the reason is maybe just that it is a “youtube star” and so not worthy
yet any consideration. (More likely is that the object is just too difficult to study as we lack
the mathematical analytic tools which for example would just to answer the basic question
whether M is connected.) A second example is catastrophe theory [335, 410] a beautiful
part of singularity theory which started with Hassler Whitney and was developed by René
Thom [404]. It was hyped to much that it fell into a fall from which it has not yet fully
recovered. And this despite the fact that Thom himself pointed out the limits as well as the
controversies of the theory already. [53] It had to pay a prize for its fame and appears to be
forgotten. Chaos theory from the 60ies which started to peak with Edward Lorenz, Mandelbrot
and strange attractors etc started to become a clishé latest after that infamous scene featuring
the character Ian Malcolm in the 1993 movie Jurassic park. It was already laughed at within
the same movie franchise, when in the third Jurassic Park installment of 2001, the kid Erik
Kirby snuffs on Malcolm’s “preachiness” and quotes his statement “everything is chaos” in a
condescending way. In art, architecture, music, fashion or design also, if something has become
too popular, it is despised by the “connaisseurs”. Hardly anybody would consider a “lava
lamp” (invented in 1963) a object of taste nowadays, even so, the fluid dynamics and motion is
objectively rich and interesting. The piano piece “Für Elise” by Ludwig van Beethoven became
so popular that it can not even be played any more as background music in a supermarket.
There is something which prevents a “music critics” to admit that the piece is great. Such
examples suggest that it might be better for an achievement (or theorem in mathematics) not
to enter pop-culture as this shows lack of “deepness” and is despised by the elite. The principle
of having fame torn down to disgrace is common also outside of mathematics. Famous actors,
entrepreneurs or politicians are not all admired but sometimes also hated to the guts, or torn
to pieces and certainly can not live normal lives any more. The phenomenon of accumulated
critique got amplified with mob type phenomena in social media. There must be something
fulfilling to trash achievements. Film critics are often harsh and judge negatively because this
elevates their own status as they appear to have a “high standard”. Similarly morale judgement
is expressed often just to elevate the status of the judge even so experience has shown that often
judges are offenders themselves. Maybe it is also human Schadenfreude, or greed which makes
so many to voice critique. History shows however that social value systems do not matter much
in the long term. A good and rich theory will show its value if it is appreciated also in hundreds
of years, where fashion and social influence has less impact. The theorem of Pythagoras will
be important independent of fame and even if it has become a cliché, it is too important to be
labeled as such. It has not only earned the status of kitsch, it is also a prototype as well as a
useful tool.
84
OLIVER KNILL
Media
There is no question that the Pythagoras theorem, the Euler polyhedron formula χ =
v − e + f the Euler identity eiπ + 1 = 0, or the Basel problem formula 1 + 1/4 + 1/9 +
1/16 + · · · = π/6 will always rank highly in any list of beautiful formulas. Most mathematicians
agree that they are elegant and beautiful. These results will also in the future keep top spots in
any ranking. On social networks, one can find lists of√favorite formulas. On “Quora”, one can
find the arithmetic mean-geometric mean inequality ab ≤ (a + b)/2 or the geometric sum-
mation formula 1 + a + a2 + · · · = 1/(1 − a) high up. One can also find strange contributions
in social media like the identity 1 = 0.99999 . . . (which is used by Piaget inspired educators
to probe mathematical maturity of kids. Similarly as in Piaget’s experiments, there is time of
mathematical maturity where a student starts to understand that this is an identity. A very
young student thinks 1 is larger than 0.9999... even if told to point out a number in between.
Such threshold moments can be crucial for example to mathematical success later. We have a
strange fascination with “wunderkinds”, kids for which some mathematical abilities have come
earlier (even so the existence of each wonder kid produces a devastating collateral damage in
its neighborhood as their success sucks out any motivation of immediate peers). The problem
is also that if somebody does not pass these Piaget thresholds early, teachers and parents con-
sider them lost, they get discouraged and become uninterested in math (the situation in other
art or sport is similar). In reality, slow learners for which the thresholds are passed later are
often deeper thinkers and can produce deeper or more extraordinary results.) At the moment,
searching for the “most beautiful formula in mathematics” gives the Euler identity and search
engines agree. But the concept of taste in a time of social media can be confusing. We live in
a time, when a 17 year old “social influencer” can in a few days gather more “followers” and
become more widely known than Sophie Kovalewskaya who made fundamental beautiful
and lasting contributions in mathematics and physics like the Cauchy-Kovalevskaya theorem
mentioned above. This theorem is definitely more lasting than a few “selfie shots” of a pretty
face, but measured by a “majority vote”, it would not only lose, but completely disappear.
One can find youtube videos of kids explaining the 4th dimension, which are watched millions
of times, many thousand times more than videos of mathematicians who have created deep
mathematical new insight about four dimensional space. But time rectifies. Kovalewskaya will
also be ranked highly in 50 years, while the pretty face has faded. Hardy put this even more
extreme by comparing a mathematician with a literary heavy weight: Archimedes will be re-
membered when Aeschylus is forgotten, because languages die and mathematical ideas do not.
[185] There is no doubt that film and TV (and now internet like “Youtube” and blogs) has a
great short-term influence on value or exposure of a mathematical field. Examples of movies
with influence are It is my turn (1980), or Antonia’s line (1995) featuring some algebraic
topology, Good will hunting (1997) in which some graph theory and Fourier theory appears,
21 from (2008) which has a scene in which the Monty Hall problem has a cameo. The man
who knew infinity displays the work of Ramanujan and promotes some combinatorics like
the theory of partitions. There are lots of movies featuring cryptology like Sneakers (1992),
Breaking the code (1996), Enigma (2001) or The imitation game (2014). For TV, math-
ematics was promoted nicely in Numb3rs (2005-2010). For more, see [333] or my own online
math in movies collection.
85
Professional opinions
Interviews with professional mathematicians can also probe the waters. In [260], Natasha Kon-
dratieva had asked a number of mathematicians: “What three mathematical formulas are the
most beautiful to you”. The formulas of Euler or the Pythagoras theorem naturally were
ranked high. Interestingly, Michael Atiyah included even a formula ”Beauty = Simplicity +
Depth”. Also other results, like the Leibniz series π/4 = 1 − 1/3 + 1/5 − 1/7 + 1/9 − . . . , the
Maxwell equations dF = 0, d∗ F = J or the Schrödinger equation 0
P∞ i~u s= (i~∇
Q + eA) su−1
2
+
2
V u, the Einstein formula E = mc or the Euler’s golden key n=1 1/n = p (1 − 1/p )
R∞ 2 √
or the Gauss identity −∞ e−x dx = π or the volume of the unit ball in R2n given as
π n /n! appeared. Gregory Margulis mentioned
√ P −n an application of the Poisson summation for-
P ˆ 2 2
= n e−n /4 or the quadratic reciprocity
P P
mula n f (n) = n f (n) which is 2 n e
law (p|q) = (−1)(p−1)/2(q−1)/2 , where (p|q) = 1 if q is a quadratic residue modulo p and −1
else. Robert Minlos gave the Gibbs formula, a Feynman-Kac formula or the Stirling
formula. Yakov Sinai mentioned the Gelfand-Naimark realization of an Abelian C ∗ alge-
bra as an algebra of continuous function Q∞or the second Plaw of thermodynamics. Anatoly
∞
Vershik gave the generating function k=0 (1 + x ) = n=0 p(n)xn for the partition func-
k
tion and the generalized Cauchy inequality between arithmetic and geometric mean. An
interesting statement of David Ruelle appears in that article who quoted Grothendieck by “
my life’s ambition as a mathematician, or rather my joy and passion, have constantly been
to discover obvious things . . . ”. Combining Grothendieck’s and Atiyah’s quote, fundamental
theorems should be “obvious, beautiful, simple and still deep”.
A recent column “Roots of unity” in the Scientific American asks mathematicians for their fa-
vorite theorem: examples are Noether’s theorem, the uniformization theorem, the Ham
Sandwich theorem, the fundamental theorem of calculus, the circumference of the
circle, the classification of compact 2-surfaces, Fermat’s little theorem, the Gromov
non-squeezing theorem, a theorem about Betti numbers, the Pythagorean theorem, the
classification of Platonic solids, the Birkhoff ergodic theorem, the Burnside lemma,
the Gauss-Bonnet theorem, Conways rational tangle theorem, Varignon’s theorem, an
upper bound on Reidemeister moves in knot theory, the asymptotic number of rela-
tive prime pairs, the Mittag Leffler theorem, a theorem about spectral sparsifiers, the
Yoneda lemma and the Brouwer fixed point theorem. These interviews illustrate also
that the choices are different if asked for “personal favorite theorem” or “objectively favorite
theorem”.
Fundamental versus important

Asking for fundamental theorems is different than asking for “deep theorems” or “important
theorems”. Examples of deep theorems are the Atiyah-Singer or Atiyah-Bott theorems in
differential topology, the KAM theorem related to the strong implicit function theorem, or
the Nash embedding theorem in Riemannian geometry. An other example is the Gauss-
Bonnet-Chern theorem in Riemannian geometry or the Pesin theorem in partially hy-
perbolic dynamical systems. Maybe the shadowing lemma in hyperbolic dynamics is more
fundamental than the much deeper Pesin theorem (which is still too complex to be proven
with full details in any classroom). One can also argue, whether the “theorema egregium”
86
OLIVER KNILL
of Gauss, stating that the curvature of a surface is intrinsic and not dependent on an em-
bedding is more “fundamental” than the “Gauss-Bonnet” result, which is definitely deeper.
In number theory, one can argue that the quadratic reciprocity formula is deeper than
the little Theorem of Fermat or the Wilson theorem. (The later gives an if and only
criterion for primality but still is far less important than the little theorem of Fermat which
as the later is used in many applications.) The last theorem of Fermat is an example of
an important theorem as it is deep and related to other fields and culture, but it is not yet
so much a “fundamental theorem”. Similarly, the Perelman theorem fixing the Poincaré
conjecture is important, but it is not (yet) a fundamental theorem. It is still a mountain peak
and not a sediment in a rock. Important theorems are not much used by other theorems as
they are located at the end of a development. Also the solution to the Kepler problem on
sphere packings or the proof of the 4-color theorem [80] or the proof of the Feigenbaum
conjectures [108, 213] are important results but not so much used by other results. Important
theorems build the roof of the building, while fundamental theorems form the foundation on
which a building can be constructed. But this can depend on time as what is the roof today,
might be in the foundation later on, once more floors have been added.
Open problems
The importance of a result is also related to related to open problems attached to the theorem.
Open problems fuel new research and new concepts. Of course this is a moving target but any
“value functional” for “fundamental theorems” is time dependent and a bit also a matter
of fashion, entertainment (TV series like “Numbers” or Hollywood movies like “good will
hunting” changed the value) and under the influence of big shot mathematicians which
serve as “influencers”. Some of the problems have prizes attached like the 23 problems
of Hilbert, the 15 problems of Simon [367], the 18 problems of Smale, or the 10
Millenium problems. There are beautiful open problems in any major field and building
a ranking would be as difficult as ranking theorems. (I personally believe that Landau’s list
of 4 problems are clearly on the top. They are shockingly short and elementary but brutally
hard, having resisted more than a century of attacks by the best minds). There are other
problems, where one believes that the mathematics has just not been developed yet to tackle
it, an example being the 3k+1 problem. There appears to be wide consensus that the Riemann
hypothesis is the most important open problem in mathematics. It states that the roots of the
Riemann zeta function are all located on the axes Re(z) = 1/2. In number theory, the prime
twin problem or the Goldbach problem have a high exposure as they can be explained
to a general audience without mathematics background. For some reason, an equally simple
problem, the Landau problem asking whether there are infinitely many primes of the form
n2 + 1 is less well known. In recent years, due to an alleged proof by Shinichi Mochizuki of
the ABC conjecture using a new theory called Inter-Universal Teichmuller Theory (IUT)
which so far is not accepted by the main mathematical community despite strong efforts. But
it has put the ABC conjecture from 1985 in the spot light like [431]. It has been described
in [163] as the most important problem in Diophantine equations. It can be expressed using
the quality Q(a, b, c) of three integers a, b, c which is Q(a, b, c) = log(c)/ log(rad(abc)), where
the radical rad(n) of a number n is the product of the distinct prime factors of n. The ABC
conjecture is that for any real number q > 1 there exist only finitely many triples (a, b, c) of
positive relatively prime integers with a + b = c for which Q(a, b, c) > q. The triple with
87
the highest quality so far is (a, b, c) = (2, 310 109, 235 ); its quality is Q = 1.6299. And then
there are entire collections of conjectures, one being the Langlands program which relates
different parts of mathematics like number theory, algebra, representation theory or algebraic
geometry. I myself can not appreciate this program because I would need first to understand it.
My personal favorite problem is the entropy problem in smooth dynamical systems theory
[229]. The Kolmogorov-Sinai entropy of a smooth dynamical system can be described using
Lyapunov exponents. For many systems like smooth convex billiards, one measures positive
entropy but can not prove it. An example is the table x4 + y 4 = 1 [220]. For ergodic theory see
[100, 109, 152, 372]
Classification results
One can also see classification theorems like the above mentioned Gelfand-Naimark realization
as mountain peaks in the landscape of mathematics. Examples of classification results are
the classification of regular or semi-regular polytopes, the classification of discrete subgroups of
a Lie group, the classification of “Lie algebras”, the classification of “von Neumann algebras”,
the “classification of finite simple groups”, the classification of Abelian groups, or the
classification of associative division algebras which by Frobenius is given either by the real
or complex or quaternion numbers. Not only in algebra, also in differential topology, one would
like to have classifications like the classification of d-dimensional manifolds. In topology, an
example result is that every Polish space is homeomorphic to some subspace of the Hilbert
cube. Related to physics is the question what “functionals” are important. Uniqueness results
help to render a functional important and fundamental. The classification of valuations of
fields is classified by Ostrowski’s theorem classifying valuations over the rational numbers
either being the absolute value or the p-adic norm. The Euler characteristic for example
can be characterized as the unique valuation on simplicial complexes which assumes the value
1 on simplices or functional which is invariant under Barycentric refinements. A theorem of
Claude Shannon [363] identifies the Shannon entropy is the unique functional on probability
spaces being compatible with additive and multiplicative operations on probability spaces and
satisfying some normalization condition.
Bounds and inequalities

An other class of important theorems are best bounds like
√ the Hurwitz estimate stating that
2
there are infinitely many p/q for which |x − p/q| < 1/( 5q ). In packing problems, one wants
to find the best packing density, like for sphere packing problems. In complex analysis, one
has the maximum principle, which assures that a harmonic function f can not have a local
maximum in its domain of definition. One can argue for including this as a fundamental theorem
as it is used by other theorems like the Schwarz lemma (named after Hermann Amandus
Schwarz) from complex analysis which is used in many places. In probability theory or statistical
mechanics, one often has thresholds, where some phase transition appears. Computing these
values is often important. The concept of maximizing entropy explains many things like
why the Gaussian distribution is fundamental as it maximizes entropy. Measures maximizing
entropy are often special and often equilibrium measures. This is a central topic in statistical
mechanics [350, 351]. In combinatorial topology, the upper bound theorem was a milestone.
It was long a conjecture of Peter McMullen and then proven by Richard Stanley that cyclic
polytopes maximize the volume in the class of polytopes with a given number of vertices.
88
OLIVER KNILL
Fundamental area also some inequalities [159] like the Cauchy-Schwarz inequality |a · b| ≤
|a||b|, the Chebyshev inequality P[|X − [E[X]| ≥ |a|] ≤ Var[X]/a2 . In complex analysis, the
Hadamard three circle theorem is important as gives bounds between the maximum of
|f | for a holomorphic function f defined on an annulus given by two concentric circles. Often
inequalities are more fundamental and powerful than equalities because they are more widely
used. Related to inequalities are embedding theorems like Sobolev embedding theorems.
For more inequalities, see [67]. Apropos embedding, there are the important Whitney or Nash
embedding theorems which are appealing.
Big ideas
Classifying and valuing big ideas is even more difficult than ranking individual theorems. Ex-
amples of big ideas are the idea of axiomatisation which stated with planar geometry and
number theory as described by Euclid and the concept of proof or later the concept of mod-
els. Archimedes idea of comparison, leading to ideas like the Cavalieri principle, integral
geometry or measure theory. René Descartes idea of coordinates which allowed to work on
geometry using algebraic tools, the use of infinitesimals and limits leading to calculus, al-
lowing to merge concepts of rate of change and accumulation, the idea of extrema leading
to the calculus of variations or Lagrangian and Hamiltonian dynamics or descriptions of fun-
damental forces. Cantor’s set theory allowed for a universal simple language to cover all
of mathematics, the Klein Erlangen program of “classifying and characterizing geometries
through symmetry”. The abstract idea of a group or more general mathematical structures
like monoids. The concept of extending number systems like completing the real numbers
or extending it to the quaternions and octonions or then producing p-adic number or
hyperreal numbers. The concept of complex numbers or more generally the idea of com-
pletion of a field. The idea of logarithms [382]. The idea of Galois to relate problems about
solving equations with field extensions and symmetries. The Grothendieck program
of “geometry without points” or “locales” as topologies without points in order to overcome
shortcomings of set theory. This lead to new objects like schemes or topoi. An other basic
big idea is the concept of duality, which appears in many places like in projective geometry,
in polyhedra, Poincaré duality or Pontryagin duality or Langlands duality for reductive
algebraic groups. The idea of dimension to measure topological spaces numerically leading
to fractal geometry. The idea of almost periodicity is an important generalization of
periodicity. Crossing the boundary of integrability leads to the important paradigm of stabil-
ity and randomness [307] and the interplay of structure and randomness [400]. These themes
are related to harmonic analysis and integrability as integrability means that for every
invariant measure one has almost periodicity. It is also related to spectral properties in solid
state physics or via Koopman theory in ergodic theory or then to fundamental new number
systems like the p-adic numbers: the p-adic integers form a compact topological group on
which the translation is almost periodic. It also leads to problems in Diophantine approxi-
mation. The concept of algorithm and building the foundation of computation using precise
mathematical notions. The use of algebra to track problems in topology starting with Betti
and Poincaré. An other important principle is to reduce a problem to a fixed point problem.
The categorical approach is not only a unifying language but also allows for generalizations
of concepts allowing to solve problems. Examples are generalizations of Lie groups in the form
of group schemes. Then there is the deformation idea which was used for example in the
Perelman proof of the Poincaré conjecture. Deformation often comes in the form of partial
89
differential equations and in particular heat type equations. Deformations can be abstract in
the form of homotopies or more concrete by analyzing concrete partial differential equations
like the mean curvature flow or Ricci flow. An other important line of ideas is to use prob-
ability theory to prove results, even in combinatorics. A probabilistic argument can often
give existence of objects which one can not even construct. Examples are graphs with n nodes
for which the Euler characteristic of the defining Whitney complex is exponentially large
in n. The idea of non-commutative geometry generalizing geometry through functional
analysis or the idea of discretization which leads to numerical methods or computational
geometry. The power of coordinates allows to solve geometric problems more easily. The above
mentioned examples have all proven their use. Grothendieck’s ideas lead to the solution of the
Weyl conjectures, fixed point theorems were used in Game theory, to prove uniqueness of
solutions of differential equations or justify perturbation theory like the KAM theorem about
the persistence of quasi-periodic motion leading to hard implicit function theorems. In the
end, what really counts is whether the big idea can solve problems or to prove theorems. The
history of mathematics clearly shows that abstraction for the sake of abstraction or for the sake
of generalization rarely could convince the mathematical community. At least not initially. But
it can also happen that the break through of a new theory or generalization only pays off much
later. A big idea might have to age like a good wine.
Paradigms
There is once in a while an idea which completely changes the way we look at things. These
are paradigm shifts as described by the philosopher and historian Thomas Kuhn who relates
it also to scientific revolutions [268]. For mathematics, there are various places, where
such fundamental changes happened: the introduction of written numbers which happened
independently in various different places. An early example is the tally mark notation on
tally sticks (early sources are the Lebombo bone from 40 thousand years ago or the Ishango
bone from 20 thousand years ago) or the technology of talking knots, the khipu [409], which
is a topological writing which flourished in the Tawantinsuyu, the Inka empire. An other
example of a paradigm change is the development of proof, which required the insight that
some mathematical statements are assumed as axioms from which, using logical deduction,
new theorems are proven. That axiom systems can be deformed like from Euclidean to non-
Euclidean geometry was definitely a paradigm change too. On a larger scale, the insight that
even the axiom systems of mathematics can be deformed and extended in various ways came
only in the 20th century with Gödel. A third example of a paradigm change is the introduction
of the concept of functions which came surprisingly late. The modern concept of a function
which takes a quantity and assigns it a new quantity came only late in the 19’th century with the
development of set theory, which is a paradigm change too. There had been a long struggle
also with understanding limits, which puzzled already Greek mathematicians like Zeno but
which really only became solid with clear definitions like Weierstrass and then with the concept
of topology where the concept of limit is absorbed within set theory, for example using filters.
Related to functions is the use of functions to understand combinatorial or number theoretical
problems, like through the use of generating functions, or Dirichlet series, allowing analytic
tools to solve discrete problems like the existence of primes on arithmetic progressions. The
opposite, the use of discrete structures like finite groups to understand the continuum like Galois
theory is an other example of a paradigm change. It led to the insight that the quadrature
90
OLIVER KNILL
of the circle, or angle trisection can not be done with ruler and compass. There are various
other places, where paradigm changes happened. A nice example is the axiomatization of
Probability theory by Kolmogorov or the realization that statistics is a geometric theory:
random variables are vectors in a vector space. The correlation between two random variables
is the cosine of the angle between centered versions of these random variables. Paradigm
changes which are really fundamental can be surprisingly simple. An example is the Connes
formula [90] which is based on the simple idea that distance can be measured by extemizing
slope. This allows to push geometry into non-commutative settings or discrete settings, where
a priory no metric is given. An other example is the extremely simple but powerful idea of the
Grothendieck extension of a monoid to a group. It has been used throughout the history
of mathematics to generate integers from natural numbers, rational numbers from integers,
complex numbers from real numbers or quaternions from complex numbers. The idea is also
used in dynamical systems theory to generate from a not necessarily invertible dynamical system
an invertible dynamical system by extending time from a monoid to a group. In the context
of Grothendieck, one should mention also that category theory similarly as set theory at the
beginning of the last century changes the way mathematics is done and extended.
Taxonomies
When looking at mathematics overall, taxonomies are important. They not only help to
navigate the landscape and are also interesting from a pedagogical as well as historical point
of view. I borrow here some material from my course Math E 320 which is so global that a
taxonomy is helpful. Organizing a field using markers is also important when teaching intelli-
gent machines, a field which be seen as the pedagogy for AI. The big bulk of work in [254]
was to teach a bot mathematics, which means to fill in thousands of entries of knowledge. It
can appear a bit mind numbing as it is a similar task than writing a dictionary. But writing
things down for a machine actually is even tougher than writing things down for a student. We
can not assume the machine to know anything it is not told. This document by the way could
relatively easily be adapted into a database of “important theorems” and actually one my aims
is it to feed it eventually to the Sofia bot. If the machine is asked about “important theorem
in mathematics”, it would be surprisingly well informed, even so it is just stupid encyclopedic
data entry. Historically, when knowledge was still sparse, one has classified teaching material
using the liberal arts of sciences, the trivium: grammar, logic and rhetoric, as well as the
quadrivium: arithmetic, geometry, music, and astronomy. More specifically, one has built the
eight ancient roots of mathematics which are tied to activities: counting and sorting (arith-
metic), spacing and distancing (geometry), positioning and locating (topology), surveying and
angulating (trigonometry), balancing and weighing (statics), moving and hitting (dynamics),
guessing and judging (probability) and collecting and ordering (algorithms). This led then to
the 12 topics taught in that course: Arithmetic, Geometry, Number Theory, Algebra, Calculus,
Set theory, Probability, Topology, Analysis, Numerics, Dynamics and Algorithms. The AMS
classification is much more refined and distinguishes 64 fields. The Bourbaki point of view
is given in [114]: it partitions mathematics into algebraic and differential topology, differential
geometry, ordinary differential equations, Ergodic theory, partial differential equations, non-
commutative harmonic analysis, automorphic forms, analytic geometry, algebraic geometry,
number theory, homological algebra, Lie groups, abstract goups, commutative harmonic analy-
sis, logic, probability theory, categories and sheaves, commutative algebra and spectral theory.
What are hot spots in mathematics? Michael Atiyah [23] distinguished parameters like
91
local - global, low and high dimensional, commutative - non-commutative, linear -

nonlinear, geometry - algebra, physics and mathematics.
Key examples
The concept of experiment came even earlier and has always been part of mathematics. Ex-
periments allow to get good examples and set the stage for a theorem. 2 Obviously the theorem
can not contradict any of the examples. But examples are more than just a tool to falsify state-
ments; a good example can be the seed for a new theory or for an entire subject. Here are
a few examples: in smooth dynamical systems the Smale horse shoe comes to mind, in
differential topology the exotic spheres of Milnor, in one-dimensional dynamics the lo-
gistic map, or Henon map, in perturbation theory of Hamiltonian systems the Standard
map featuring KAM tori or Mather sets, in homotopy theory the dunce hat or Bing house,
in combinatorial topology the Rudin sphere, the Nash-Kuiper non-smooth embedding
of a torus into Euclidean space, in topology there is the Alexander horned sphere or the
Antoine necklace. In complexity theory there is the busy beaver problem in Turing com-
putation which is an illustration with how small machines one can achieve great things, in
group theory there is the Rubik cube which illustrates many fundamental notions for finite
groups, in fractal analysis the Cantor set, the Menger sponge, in Fourier theory the series
of f (x) = x mod 1, in Diophantine approximation the golden ratio, in the calculus of sums
the zeta function, in dimension theory the Banach Tarski paradox. In harmonic analysis
the Weierstrass function as an example of a nowhere differentiable function. The case of
Peano curves giving concrete examples of a continuous bijection from an interval to a square
or cube. In complex dynamics not only the Mandelbrot set plays an important role, but
also individual, specific Julia sets can be interesting. Examples like the Mandelbulb have not
even been started to be investigated. There seem to be no theorems known about this object.
In mathematical physics, the almost Matthieu operator [105] produced a rich theory related
to spectral theory, Diophantine approximation, fractal geometry and functional analysis.
Besides examples illustrating a typical case, it is also important to explore the boundary of a
theorem or theory by looking at counter examples. Collections of counter examples exist in
many fields like [158, 383, 338, 392, 429, 72, 243].
Physics
One can also make a list of great ideas in physics [126] and see the relations with the fun-
damental theorems in mathematics. A high applicability should then contribute to a value
functional in the list of theorems. Great ideas in physics are the concept of space and time
meaning to describe physical events using differential equations. In cosmology, one of the
insights is the to understand the structure of our solar system and getting to the heliocentric
system, an other is to look at space-time as a hole and realize the expansion of the universe
or that the idea of a big bang. More general is the Platonic idea that physics is geometry.
Or calculus: Lagrange developed his calculus of variations to find laws of physics. Then
there is the idea of Lorentz invariance which leads to special relativity, there is the idea of
general relativity which allows to describe gravity through geometry and a larger symmetry
seen through the equivalence principle. There is the idea of see elementary particles using
Lie groups. There is the Noether theorem which is the idea that any symmetry is tied
to a conservation law: translational symmetry leads to momentum conservation, rotation
2Quote of Vladimir Arnold: ”Mathematics is a part of physics where experiments are cheap”
92
OLIVER KNILL
symmetry to angular momentum conservation for example. Symmetries also play a role when
spontaneous broken symmetry or phase transitions. There is the idea of quantum me-
chanics which mathematically means replacing differential equations with partial differential
equations or replacing commutative algebras of observables with non-commutative alge-
bras. An important idea is the concept of perturbation theory and in particular the notion
of linearization. Many laws are simplifications of more complicated laws and described in
the simplest cases through linear laws like Ohms law or Hooks law. Quantization processes
allow to go from commutative to non-commutative structures. Perturbation theory allows
then to extrapolate from a simple law to a more complicated law. Some is easy application
of the implicit function theorem, some is harder like KAM theory. There is the idea of
using discrete mathematics to describe complicated processes. An example is the language
of Feynman graphs or the language of graph theory in general to describe physics as in loop
quantum gravity or then the language of cellular automata which can be seen as partial
difference equations where also the function space is quantized. The idea of quantization, a
formal transition from an ordinary differential equation like a Hamiltonian system to a partial
differential equation or to replace single particle systems with infinite particle systems (Fock).
There are other quantization approaches through deformation of algebras which is related
to non-commutative geometry. There is the idea of using smooth functions to describe
discrete particle processes. An example is the Vlasov dynamical system or Boltzmann’s
equation to describe a plasma, or thermodynamic notions to describe large sets of particles like
a gas or fluid. Dual to this is the use of discretization to describe a smooth system by discrete
processes. An example is numerical approximation, like using the Runge-Kutta scheme to
compute the trajectory of a differential equation. There is the realization that we have a whole
spectrum of dynamical systems, integrability and chaos and that some of the transitions are
universal. An other example is the tight binding approximation in which a continuum
Schrödinger equation is replaced with a bounded discrete Jacobi operator. There is the gen-
eral idea of finding the building blocks or elementary particles. Starting with Demokrit in
ancient Greece, the idea got refined again and again. Once, atoms were detected and charges
found to be quantized (Millikan), the structure of the atom was explored (Rutherford), and
then the atom got split (Meitner, Hahn). The structure of the nuclei with protons and neutrons
was then refined again using quarks leading the standard model in particle physics. There
is furthermore the idea to use statistical methods for complex systems. An example is the
use of stochastic differential equations like diffusion processes to describe actually deterministic
particle systems. There is the insight that complicated systems can form patterns through in-
terplay between symmetry, conservation laws and synchronization. Large scale patterns can
be formed from systems with local laws. Finally, there is the idea of solving inverse problems
using mathematical tools like Fourier theory or basic geometry (Erathostenes could compute
the radius of the earth by comparing the lengths of shadows at different places of the earth.)
An example is tomography, where the structure of some object is explored using resonance.
Then there is the idea of scale invariance which allows to describe objects which have fractal
nature.
Computer science
As in physics, it is harder to pinpoint “big ideas” in computer science as they are in general
not theorems. The initial steps of mathematics was to build a language, where numbers
93
represent quantities [97]. Physical tools which assist in manipulating numbers can already
been seen as a computing device. Marks on a bone, pebbles in a clay bag, talking knots in a
Khipu, marks on a Clay tablet were the first step. Papyri, paper, magnetic, optical and electric
storage, the tools to build memory were refined over the millenniums. The mathematical
language allowed us to get further than the finite. Using a finite number of symbols we can
represent and count infinite sets, have notions of cardinality, have various number systems
and more general algebraic structures. Numbers can even be seen as games [96, 257]. A
major idea is the concept of an algorithm. Adding or multiplying on an abacus already was
an algorithm. The concept was refined in geometry, where ruler and compass were used as
computing devices, like the construction of points in a triangle. To measure the effectiveness
of an algorithm, one can use notions of complexity. This has been made precise by computing
pioneers like Turing as one has to formulate first what a computation is. In the last century
one has seen that computations and proofs are very similar and that they have similar general
restrictions. There are some tasks which can not be computed with a Turing machine and
there are theorems which can not be proven in a specific axiom system. As mathematics is a
language, we have to deal with concepts of syntax, grammar, notation, context, parsing,
validation, verification. As Mathematics is a human activity which is done in our brains,
it is related to psychology and computer architecture. Computer science aspects are also
important also in pedagogy and education how can an idea be communicated clearly? How
do we motivate? How do we convince peers that a result is true? Examples from history
show that this is often done by authority and that the validity of some proofs turned out
to be wrong or incomplete, even in the case of fundamental theorems or when treated by
great mathematicians. (Examples are the fundamental theorem of arithmetic, the fundamental
theorem of algebra or the wrong published proof of Kempe of the 4 color theorem). On the
other hand, there were also quite many results which only later got recognized. The work
of Galois for example only exploded much later. How come we trust a human brain more
than an electronic one? We have to make some fundamental assumptions for example to be
made like that if we do a logical step ”if A and B then “A and B” holds. This assumes for
example that our memory is faithful: after having put A and B in the memory and making
the conclusion, we have to assume that we did not forget A nor B! Why do we trust this
more than the memory of a machine? As we are also assisted more and more by electronic
devices, the question of the validity of computer assisted proofs comes up. The 4-color
theorem of Kenneth Appel and Wolfgang Haken based on previous work of many others
like Heinrich Heesch or the proof of the Feigenbaum conjecture of Mitchell Feigenbaum
first proven by Oscar Lanford III or the proof of the Kepler problem by Thomas Hales are
examples. A great general idea is related to the representation of data. This can be done using
matrices like in a relational database or using other structures like graphs leading to graph
databases. The ability to use computers allows mathematicians to do experiments. A branch
of mathematics called experimental mathematics [20, 216] relies heavily on experiments to
find new theorems or relations. Experiments are related to simulations. We are able, within
a computer to build and explore new worlds, like in computer games, we can enhance the
physical world using virtual reality or augmented reality or then capturing a world by
3D scanning and realize a world by printing the objects [256]. A major theme is artificial
intelligence [354, 217]. It is related to optimization problems like optimal transport, neural
nets as well as inverse problems like structure from motion problems. An intelligent
entity must be able to take information, build a model and then find an optimal strategy to
solve a given task. A self-driving car for example has to be able to translate pictures from a
94
OLIVER KNILL
camera and build a map, then determine where to drive. Such tasks are usually considered
part of applied mathematics but they are very much related with pure mathematics because
computers also start to learn how to read mathematics, how to verify proofs and to find new
theorems. Artificial intelligent agents [421] were first developed in the 1960ies learned also
some mathematics. I myself learned about it when incorporated computer algebra systems into
a chatbots in [254]. AI has now become a big business as Alexa, Siri, Google Home, IBM
Watson or Cortana demonstrate. But these information systems must be taught, they must
be able to rank alternative answers, even inject some humor or opinions. Soon, they will be
able to learn themselves and answer questions like “what are the 10 most important theorems
in mathematics?”
Brevity
We live in a instagram, snapchat, twitter, microblog, vine, watch-mojo, petcha-kutcha time.
Many of us multi task, read news on smart phones, watch faster paced movies, read shorter
novels and feel that a million word Marcel Proust’s masterpiece “a la recherche du temps perdu”
is ”temps perdu”. Even classrooms and seminars have become more aphoristic. Micro blogging
tools are only the latest incarnation of “miniature stories”. They continue the tradition of older
formats like ”mural art” by Romans to modern graffiti or “aphorisms” [263, 264]), poetry,
cartoons, Unix fortune cookies [17]. Shortness has appeal: aphorisms, poems, ferry tales,
quotes, words of wisdom, life hacker lists, and tabloid top 10 lists illustrate this. And then there
are books like “Math in 5 minutes”, “30 second math”, “math in minutes” [35, 162, 129], which
are great coffee table additions. Also short proofs are appealing like “Let epsilon be smaller
than zero” which is the shortest known math joke, or “There are three type of mathematicians,
the ones who can count, and the ones who can’t.” Also short open problems are attractive, like
the twin prime problem “there are infinitely many twin primes” or the Landau problem
“there are infinitely many primes of the form n2 + 1, or the Goldbach problem “every n > 2 is
the sum of two primes”. For the larger public in mathematics shortness has appeal: according
to a poll of the Mathematical Intelligencer from 1988, the most favorite theorems are short
[423]. Results with longer proofs can make it to graduate courses or specialized textbooks but
still then, the results are often short enough so that they can be tweeted without proof. Why
is shortness attractive? Paul Erdös expressed short elegant proofs as “proofs from the book”
[9]. Shortness reduces the possibility of error as complexity is always a stumbling block for
understanding. But is beauty equivalent to brevity? Buckminster Fuller once said: “If the
solution is not beautiful, I know it is wrong.” [7]. Much about the aesthetics in mathematics
is investigated in [304]. According to [348], the beauty of a piece of mathematics is frequently
associated with the shortness of statement or of proof: beautiful theories are also thought of
as short, self-contained chapters fitting within broader theories. There are examples of complex
and extensive theories which every mathematician agrees to be beautiful, but these examples
are not the one which come to mind. Also psychologists and educators know that simplicity
appeals to children: From [373] For now, I want simply to draw attention to the fact that
even for a young, mathematically naive child, aesthetic sensibilities and values (a penchant for
simplicity, for finding the building blocks of more complex ideas, and a preference for shortcuts
and “liberating” tricks rather than cumbersome recipes) animates mathematical experience. It
is hard to exhaust them all, even not with tweets: there are more than googool2 = 10200 texts of
length 140. This can not all ever be written down because there are more than what we estimate
the number of elementary particles. But there are even short story collections. Berry’s paradox
95
tells in this context that the shortest non-tweetable text in 140 characters can be tweeted: ”The
shortest non-tweetable text”. Since we insist on giving proofs, we have to cut corners. Books
containing lots of elegant examples are [14, 9].
Twitter math
The following 42 tweets were written in 2014, when twitter had still a 140 character limit. Some
of them were actually tweeted. The experiment was to see which theorems are short enough
so that one can tweet both the theorem as well as the proof in 140 characters. Of course, that
often requires a bit of cheating. See [9] for proofs from the books, where the proofs have full
details.
Euclid: The set of primes is infinite. Proof: let p be largest
prime, then p! + 1 has a larger prime factor than p. Contradic-
tion.
Euclid: 2p −1 prime then 2p−1 (2p −1) is perfect. Proof. σ(n) =

sum of factors of n, σ(2n − 1)2n−1 ) = σ(2n − 1)σ(2n−1 ) =
2n (2n − 1) = 2 · 2n (2n − 1) shows σ(k) = 2k.
√ √
Hippasus: 2 is irrational. Proof. If 2 = p/q, then 2q 2 =
p2 . To the left is an odd number of factors 2, to the right it is
even one. Contradiction.
Pythagorean triples: all x2 + y 2 = z 2 are of form (x, y, z) =

(2st, s2 − t2 , s2 + t2 ). Proof: x or y is even (both odd gives
x2 + y 2 = wk with odd k). Say x2 is even: write x2 = z 2 − y 2 =
(z − y)(z + y). This is 4s2 t2 . Therefore 2s2 = z − y, 2t2 = z + y.
Solve for z, y.
Pigeon principle: if n + 1 pigeons live in n boxes, there is a

box with 2 or more pigeons. Proof: place a pigeon in each box
until every box is filled. The pigeon left must have a roommate.
Angle sum in triangle: α + β + γ = KA + π if K is cur-

vature, A triangle area. Proof: Gauss-Bonnet for surface with
boundary. α, β, γ are Dirac measures on the boundary.
Chinese remainder theorem: a(i) x = b(i) mod n(i) has a

solution if gcd(a(i),n(i))=0 and gcd(n(i),n(j))=0 Proof: solve
eq(1), then increment x by n(1) to solve eq(2), then increment
x by n(1) n(2) until second is ok. etc.
96
OLIVER KNILL
Nullstellensatz: algebraic sets in K n are 1:1 to radical ideals

in K[x1 ...xn ]. Proof: An algebra over K which is a field is finite
field extension of K.
Fundamental theorem algebra: a polynomial of degree n

has exactly n roots. Proof: the metric g = |f |−2/n |dz|2 on
the Riemann sphere has curvature K = n−1 ∆ log |f |. Without
root, K=0 everywhere contradicting Gauss-Bonnet. [12]:
Fermat: p prime (a, p) = 1, then p|ap − a Proof: induction

with respect to a. Case a = 1 is trivial (a + 1)p − (a + 1)
is congruent to ap − a modulo p because Binomial coefficients
B(p, k) are divisible by p for k = 1, . . . p − 1.
Wilson: p is prime iff p|(p − 1)! + 1 Proof. Group 2, . . . p − 2

into pairs (a, a− 1) whose product is 1 modulo p. Now (p−1)! =
(p−1) = −1 modulo p. If p = ab is not prime, then (p−1)! = 0
modulo p and p does not divide (p − 1)! + 1.
Bayes: A, B are events and Ac is the complement. P [A|B] =

P [B|A]P [A]/(P [B|A]P [A] + P [B|Ac ]P [Ac ] Proof: By defini-
tion P [A|B]P [B] = P [A ∩ B]. Also P [B] = (P [B|A]P [A] +
P [B|Ac ]P [Ac ].
Archimedes: Volume of sphere S(r) is 4πr3 /3 Proof: the

complement of the cone inside the cylinder has at height z the
cross section area r2 − z 2 , the same as the cross section area of
the sphere at height z.
Archimedes: the area of the sphere S(r) is 4πr2 Proof: dif-

ferentiate the volume formula with respect to r or project the
sphere onto a cylinder of height 2 and circumference 2π and
not that this is area preserving.
Cauchy-Schwarz: |v·w| ≤ |v||w|. Proof: scale to get |w| = 1,

define a = v.w, so that 0 ≤ (v − aw) . . . (v − aw) = |v|2 − a2 =
|v|2 |w|2 − (v · w)2 .
Angle formula: Cauchy-Schwarz defines the angle between

two vectors as cos(A) = v.w/|v||w|. If v, w are centered random
variables, then v · w is the covariance, |v|, |w| are standard
deviations and cos(A) is the correlation.
97
Cos formula: c2 = a2 + b2 − ab cos(A) in a triangle ABC

(Al-Kashi theorem) Proof: v = AB, w = AC has length a =
|v|, b = |w|, |c| = |v − w|. Now: (v − w).(v − w) = |v|2 + |w|2 −
2|v||w| cos(A).
Pythagoras: A = π/2, then c2 = a2 + b2 . Proof: Let v = AB,

w = AC, v − w = BC be the sides of the triangle. Multiply
out (v − w) · (v − w) = |v|2 + |w|2 and use v · w = 0.
Euler formula: exp(ix) = cos(x) + i sin(x). Proof: exp(ix) =

1 + (ix) + (ix)2 /2! − ... Pair real and imaginary parts and use
definition cos(x) = 1 − x2 /2! + x4 /4!... and sin(x) = x − x3 /3! +
x5 /5! − .....
P
Discrete Gauss-Bonnet x K(x) = χ(G) with K(x) = 1 −
0 − v1 + v2 −
V0 (x)/2 + V1 (x)/3 + V2 (x)/4... curvature χ(G) = vP
v3 ... Euler characteristic Proof: Use handshake x Vk (x) =
vk+1 /(k + 2).
Poincaré-Hopf: let f be a coloring, if (x) = 1 − χ(Sf− (x)),

where Sf− (x) = y ∈ S(x)|f (y) < f (x)
P
if (x) = χ(G). Proof
by induction. Removing local maximum of f reduces Euler
characteristic by χ(Bf (x)) − χ(S − f (x)) = if (x).
P
Lefschetz: x iT (x) = str(T |H(G)). Proof: LHS is
str(exp(−0L)UT ) and RHS is str(exp(−tL)UT ) for t → ∞.
The super trace does not depend on t.
Stokes: orient edges E of graph G. F : E → R function, S

surface in G with boundary C. d(F )(ijk) = F (ij) + F (jk) −
F (ki) is the curl. The sum of the curls over all triangles is the
line integral of F along C.
Plato: there are exactly 5 platonic solids. Proof: number

f of n-gon satisfies f = 2e/n, v vertices of degree m satisfy
v = 2e/m v−e+f −2 means 2e/m−e+2e/n = 2 or 1/m+1/n =
1/e + 1/2 with solutions: (m = 4, n = 3), (m = 3, n = 5), (n =
m = 3), (n = 3, m = 5), (m = 3, n = 4).
98
OLIVER KNILL
Poincaré recurrence: T area-preserving map of probabil-

ity space (X, m). If m(A) > 0 and n > 1/m(A) we have
m(T k (A) ∩ A) > 0 for some 1 ≤ k ≤ n Proof. Otherwise
A, T (A), ..., T n (A) are all disjoint and the union has measure
n · m(A) > 1.
Turing: there is no Turing machine which halts if input is

Turing machine which halts: Proof: otherwise build an other
one which halts if the input is a non-halting one and does not
halt if input is a halting one.
Cantor: the set of reals in [0,1] is uncountable. Proof: if there

is an enumeration x(k), let x(k, l) be the l’th digit of x(k)
in binary form. The number with binary expansion y(k) =
x(k, k) + 1 mod 2 is not in the list.
Niven: π ∈ / Q: Proof: π = a/b, f (x) = xn (a−bx)n /n! satisfies

f (pi − x) = f (x) and 0 < f (x) < π n an /nn f ( j)(x) = 0 at 0 and
π for 0 ≤ j ≤ n shows F (x) = f (x) − f (2) (x) + f (4) (x) · · · +
(−1)n f (2n) (x) has F (0), F (π) ∈ Z and F R+ F 00 = f . Now
π
(F 0 (x) sin(x) − F (x) cos(x)) = f sin(x), so 0 f (x) sin(x)dx ∈
Z.
Fundamental theorem calculus: With differentiation

Df (x) = f (x+1)−f (x) and integration Sf (x) = f (0)+f (1)+
... + f (n − 1) have SDf (x) = f (x) − f (0), DSf (x) = f (x).
Taylor: f (x + t) = k f (k) (x)tk /k!. Proof: f (x + t) satisfies

P
transport equation ft = fx = Df an ODE for the differential
operator D. Solve f (x + t) = exp(Dt)f (x).
Cauchy-Binet: det(1 + F T G) =
P
P det(FP ) det(GP )
T
Proof: A = F G.
P Coefficients of det(x − A) is
|P |=k det(F P ) det(G P ).
Intermediate: f continuous f (0) < 0, f (1) > 0, then there

exists 0 < x < 1, f (x) = 0. Proof. If f (1/2) < 0 do proof with
(1/2, 1) If f (1/2) > 0 redo proof with (0, 1/2).
Ergodicity: T (x) = x + a mod P 1 with irrational a

is
P ergodic. Proof. f = n a(n) exp(inx) T f =
n a(n) exp(ina) exp(inx) = f implies a(n) = 0.
99
Benford: first digit k of 2n appears with probability log(1 −

1/k) Proof: T : x → x + log(2) mod 1 is ergodic.
log(2n )mod 1 = k if log(k) ≤ T n (0) < log(k + 1). The proba-
bility of hitting this interval is log(k + 1)/ log(k).
Rank-Nullity: dim(ker(A)) + dim(im(A)) = n for m × n

matrix A. Proof: a column has a leading 1 in rref (A) or no
leading 1. In the first case it contributes to the image, in the
second to a free variable parametrizing the kernel.
Column-Row picture: A : Rm → Rn . The k’th column of A

is the image Aek . If all rows of A are perpendicular to x then
x is in the kernel of A.
Picard: x0 = f (x), x(0) = x0 has locally

Rt a unique solution if
1
f ∈ C . Proof: the map T (y) = 0 f (y(s)) ds is a contrac-
tion on C([0, a]) for small enough a > 0. Banach fixed point
theorem.
Banach: a contraction d(T (x), T (y)) ≤ ad(x, y) on complete

(X, d) has a unique fixed point. Proof: d(xk , xn ) ≤ ak /(1 − a)
using triangle inequality and geometric series. Have Cauchy
sequence.
Liouville: every prime p=4k+1 is the sum of two squares.

Proof: there is an involution on S = (x, y, z)|x2 + 4yz = p
with exactly one fixed point showing —S— is odd implying
(x, y, z)− > (x, z, y) has a fixed point. [436]
Banach-Tarski: The unit ball in R3 can be cut into 5 pieces,

re-assembled using rotation and translation to get two spheres.
Proof: cut cleverly using axiom of choice.
Math areas
We add here the core handouts of Math E320 which aimed to give for each of the 12 math-
ematical subjects an overview on two pages. For that course, I had recommended books like
[139, 170, 42, 390, 391].
100
OLIVER KNILL
E-320: Teaching Math with a Historical Perspective O. Knill, 2010-2018
Lecture 1: Mathematical roots

Similarly, as one has distinguished the canons of rhetorics: memory, invention, delivery, style, and arrange-
ment, or combined the trivium: grammar, logic and rhetorics, with the quadrivium: arithmetic, geometry,
music, and astronomy, to obtain the seven liberal arts and sciences, one has tried to organize all mathe-
matical activities.
counting and sorting arithmetic
spacing and distancing geometry
Historically, one has distin-
positioning and locating topology
guished eight ancient roots of
surveying and angulating trigonometry
mathematics. Each of these 8
balancing and weighing statics
activities in turn suggest a key
moving and hitting dynamics
area in mathematics:
guessing and judging probability
collecting and ordering algorithms
To morph these 8 roots to the 12 mathematical areas covered in this class, we complemented the ancient roots
with calculus, numerics and computer science, merge trigonometry with geometry, separate arithmetic into
number theory, algebra and arithmetic and turn statics into analysis.
counting and sorting arithmetic
spacing and distancing geometry
positioning and locating topology
dividing and comparing number theory
Let us call this modern adapta-
balancing and weighing analysis
tion the
moving and hitting dynamics
12 modern roots of guessing and judging probability
Mathematics: collecting and ordering algorithms
slicing and stacking calculus
operating and memorizing computer science
optimizing and planning numerics
manipulating and solving algebra
Arithmetic numbers and number systems

Geometry invariance, symmetries, measurement, maps
While relating mathe- Number theory Diophantine equations, factorizations
matical areas with hu- Algebra algebraic and discrete structures
man activities is useful, Calculus limits, derivatives, integrals
it makes sense to select Set Theory set theory, foundations and formalisms
specific topics in each of Probability combinatorics, measure theory and statistics
this area. These 12 top- Topology polyhedra, topological spaces, manifolds
ics will be the 12 lectures Analysis extrema, estimates, variation, measure
of this course. Numerics numerical schemes, codes, cryptology
Dynamics differential equations, maps
Algorithms computer science, artificial intelligence
Like any classification, this chosen division is rather arbitrary and a matter of personal preferences. The 2010
AMS classification distinguishes 64 areas of mathematics. Many of the just defined main areas are broken
101
off into even finer pieces. Additionally, there are fields which relate with other areas of science, like economics,
biology or physics:a
00 General 45 Integral equations
01 History and biography 46 Functional analysis
03 Mathematical logic and foundations 47 Operator theory
05 Combinatorics 49 Calculus of variations, optimization
06 Lattices, ordered algebraic structures 51 Geometry
08 General algebraic systems 52 Convex and discrete geometry
11 Number theory 53 Differential geometry
12 Field theory and polynomials 54 General topology
13 Commutative rings and algebras 55 Algebraic topology
14 Algebraic geometry 57 Manifolds and cell complexes
15 Linear/multi-linear algebra; matrix theory 58 Global analysis, analysis on manifolds
16 Associative rings and algebras 60 Probability theory and stochastic processes
17 Non-associative rings and algebras 62 Statistics
18 Category theory, homological algebra 65 Numerical analysis
19 K-theory 68 Computer science
20 Group theory and generalizations 70 Mechanics of particles and systems
22 Topological groups, Lie groups 74 Mechanics of deformable solids

26 Real functions 76 Fluid mechanics
28 Measure and integration 78 Optics, electromagnetic theory
30 Functions of a complex variable 80 Classical thermodynamics, heat transfer
31 Potential theory 81 Quantum theory
32 Several complex variables, analytic spaces 82 Statistical mechanics, structure of matter
33 Special functions 83 Relativity and gravitational theory
34 Ordinary differential equations 85 Astronomy and astrophysics
35 Partial differential equations 86 Geophysics
37 Dynamical systems and ergodic theory 90 Operations research, math. programming
39 Difference and functional equations 91 Game theory, Economics Social and Behavioral Sciences
40 Sequences, series, summability 92 Biology and other natural sciences
41 Approximations and expansions 93 Systems theory and control
42 Fourier analysis 94 Information and communication, circuits
43 Abstract harmonic analysis 97 Mathematics education
44 Integral transforms, operational calculus
What are
local and global
low and high dimension
fancy developments
commutative and non-commutative
in mathematics today? Michael Atiyah linear and nonlinear
[23] identified in the year 2000 the geometry and algebra
following six hot spots: physics and mathematics
Also this choice is of course highly personal. One can easily add 12 other polarizing quantities which help to
distinguish or parametrize different parts of mathematical areas, especially the ambivalent pairs which produce
a captivating gradient:
regularity and randomness discrete and continuous
integrable and non-integrable existence and construction
invariants and perturbations finite dim and infinite dimensional
experimental and deductive topological and differential geometric
polynomial and exponential practical and theoretical
applied and abstract axiomatic and case based
The goal is to illustrate some of these structures from a historical point of view and show that “Mathematics is
the science of structure”.
102
OLIVER KNILL
E-320: Teaching Math with a Historical Perspective Oliver Knill, 2010-2018
Lecture 2: Arithmetic
The oldest mathematical discipline is arithmetic. It is the theory of the construction and manipulation of
numbers. The earliest steps were done by Babylonian, Egyptian, Chinese, Indian and Greek thinkers.
Building up the number system starts with the natural numbers 1, 2, 3, 4... which can be added and multiplied.
Addition is natural: join 3 sticks to 5 sticks to get 8 sticks. Multiplication ∗ is more subtle: 3 ∗ 4 means to take 3
copies of 4 and get 4 + 4 + 4 = 12 while 4 ∗ 3 means to take 4 copies of 3 to get 3 + 3 + 3 + 3 = 12. The first factor
counts the number of operations while the second factor counts the objects. To motivate 3 ∗ 4 = 4 ∗ 3, spacial
insight motivates to arrange the 12 objects in a rectangle. This commutativity axiom will be carried over to
larger number systems. Realizing an addition and multiplicative structure on the natural numbers requires to
define 0 and 1. It leads naturally to more general numbers. There are two major motivations to to build new
numbers: we want to
1. invert operations and still get results. 2. solve equations.

To find an additive inverse of 3 means solving x + 3 = 0. The answer is a negative number. To solve x ∗ 3 = 1,
we get to a rational number x = 1/3. To solve x2 = 2 one need to escape to real numbers. To solve x2 = −2
requires complex numbers.
Numbers Operation to complete Examples of equations to solve

Natural numbers addition and multiplication 5+x=9
Positive fractions addition and division 5x = 8
Integers subtraction 5+x=3
Rational numbers division 3x = 5
Algebraic numbers taking positive roots x2 = 2 , 2x + x2 − x3 = 2
Real numbers taking limits x = 1 − 1/3 + 1/5 − +...,cos(x) = x
Complex numbers take any roots x2 = −2
Surreal numbers transfinite limits x2 = ω, 1/x = ω
Surreal complex any operation x2 + 1 = −ω
The development and history of arithmetic can be summarized as follows: humans started with natural numbers,
dealt with positive fractions, reluctantly introduced negative numbers and zero to get the integers, struggled to
“realize” real numbers, were scared to introduce complex numbers, hardly accepted surreal numbers and most
do not even know about surreal complex numbers. Ironically, as simple but impossibly difficult questions in
number theory show, the modern point of view is the opposite to Kronecker’s ”God made the integers; all
else is the work of man”:
The surreal complex numbers are the most natural numbers;
The natural numbers are the most complex, surreal numbers.
Natural numbers. Counting can be realized by sticks, bones, quipu knots, pebbles or wampum knots. The
tally stick concept is still used when playing card games: where bundles of fives are formed, maybe by crossing
4 ”sticks” with a fifth. There is a ”log counting” method in which graphs are used and vertices and edges count.
An old stone age tally stick, the wolf radius bone contains 55 notches, with 5 groups of 5. It is probably more
than 30’000 years old. [378] The most famous paleolithic tally stick is the Ishango bone, the fibula of a baboon.
It could be 20’000 - 30’000 years old. [139] Earlier counting could have been done by assembling pebbles,
tying knots in a string, making scratches in dirt or bark but no such traces have survived the thousands of
years. The Roman system improved the tally stick concept by introducing new symbols for larger numbers
like V = 5, X = 10, L = 40, C = 100, D = 500, M = 1000. in order to avoid bundling too many single sticks.
103
The system is unfit for computations as simple calculations V III + V II = XV show. Clay tablets, some as
early as 2000 BC and others from 600 - 300 BC are known. They feature Akkadian arithmetic using the base
60. The hexadecimal system with base 60 is convenient because of many factors. It survived: we use 60 minutes
per hour. The Egyptians used the base 10. The most important source on Egyptian mathematics is the
Rhind Papyrus of 1650 BC. It was found in 1858 [234, 378]. Hieratic numerals were used to write on papyrus
from 2500 BC on. Egyptian numerals are hieroglyphics. Found in carvings on tombs and monuments they
are 5000 years old. The modern way to write numbers like 2018 is the Hindu-Arab system which diffused
to the West only during the late Middle ages. It replaced the more primitive Roman system. [378] Greek
arithmetic used a number system with no place values: 9 Greek letters for 1, 2, . . . 9, nine for 10, 20, . . . , 90 and
nine for 100, 200, . . . , 900.
Integers. Indian Mathematics morphed the place-value system into a modern method of writing numbers.
Hindu astronomers used words to represent digits, but the numbers would be written in the opposite order.
Independently, also the Mayans developed the concept of 0 in a number system using base 20. Sometimes
after 500, the Hindus changed to a digital notation which included the symbol 0. Negative numbers were
introduced around 100 BC in the Chinese text ”Nine Chapters on the Mathematical art”. Also the Bakshali
manuscript, written around 300 AD subtracts numbers carried out additions with negative numbers, where +
was used to indicate a negative sign. [330] In Europe, negative numbers were avoided until the 15’th century.
Fractions: Babylonians could handle fractions. The Egyptians also used fractions, but wrote every frac-
tion a as a sum of fractions with unit numerator and distinct denominators, like 4/5 = 1/2 + 1/4 + 1/20 or
5/6 = 1/2 + 1/3. Maybe because of such cumbersome computation techniques, Egyptian mathematics failed to
progress beyond a primitive stage. [378]. The modern decimal fractions used nowadays for numerical calcula-
tions were adopted only in 1595 in Europe.
Real numbers: As noted by the Greeks already, the diagonal of the square is not a fraction. It first produced a
crisis until it became clear that ”most” numbers are not rational. Georg Cantor saw first that the cardinality
of all real numbers is much larger than the cardinality of the integers: while one can count all rational numbers
but not enumerate all real numbers. One consequence is that most real numbers are transcendental: they do
not occur as solutions of polynomial equations with integer coefficients. The number π is an example. The
concept of real numbers is related to the concept of limit. Sums like 1 + 1/4 + 1/9 + 1/16 + 1/25 + . . . are
not rational.
Complex numbers: some polynomials have no real root. To solve x2 = −1 for example, we need new
numbers. One idea is to use pairs of numbers (a, b) where (a, 0) = a are the usual numbers and extend addition
and multiplication (a, b) + (c, d) = (a + c, b + d) and (a, b) · (c, d) = (ac − bd, ad + bc). With this multiplication,
the number (0, 1) has the property that (0, 1) · (0, 1) = (−1, 0) = −1. It is more convenient to write a + ib where
i = (0, 1) satisfies i2 = −1. One can now use the common rules of addition and multiplication.
Surreal numbers: Similarly as real numbers fill in the gaps between the integers, the surreal numbers fill in the
gaps between Cantors ordinal numbers. They are written as (a, b, c, ...|d, e, f, ...) meaning that the ”simplest”
number is larger than a, b, c... and smaller than d, e, f, ... We have (|) = 0, (0|) = 1, (1|) = 2 and (0|1) = 1/2
or (|0) = −1. Surreals contain already transfinite numbers like (0, 1, 2, 3...|) or infinitesimal numbers like
(0|1/2, 1/3, 1/4, 1/5, ...). They were introduced in the 1970’ies by John Conway. The late appearance confirms
the pedagogical principle: late human discovery manifests in increased difficulty to teach it.
104
OLIVER KNILL
Lecture 3: Geometry
Geometry is the science of shape, size and symmetry. While arithmetic deals with numerical structures,
geometry handles metric structures. Geometry is one of the oldest mathematical disciplines. Early geometry
has relations with arithmetic: the multiplication of two numbers n × m as an area of a shape that is invariant
under rotational symmetry. Identities like the Pythagorean triples 32 + 42 = 52 were interpreted and
drawn geometrically. The right angle is the most ”symmetric” angle apart from 0. Symmetry manifests
itself in quantities which are invariant. Invariants are one the most central aspects of geometry. Felix Klein’s
Erlangen program uses symmetry to classify geometries depending on how large the symmetries of the shapes
are. In this lecture, we look at a few results which can all be stated in terms of invariants. In the presentation
as well as the worksheet part of this lecture, we will work us through smaller miracles like special points in
triangles as well as a couple of gems: Pythagoras, Thales,Hippocrates, Feuerbach, Pappus, Morley,
Butterfly which illustrate the importance of symmetry.
Much of geometry is based on our ability to measure length, the distance between two points. Having a
distance d(A, B) between any two points A, B, we can look at the next more complicated object, which is a set
A, B, C of 3 points, a triangle. Given an arbitrary triangle ABC, are there relations between the 3 possible
distances a = d(B, C), b = d(A, C), c = d(A, B)? If we fix the scale by c = 1, then a + b ≥ 1, a + 1 ≥ b, b + 1 ≥ a.
For any pair of (a, b) in this region, there is a triangle. After an identification, we get an abstract space, which
represent all triangles uniquely up to similarity. Mathematicians call this an example of a moduli space.
A sphere Sr (x) is the set of points which have distance r from a given point x. In the plane, the sphere is called
a circle. A natural problem is to find the circumference L = 2π of a unit circle, or the area A = π of a unit disc,
the area F = 4π of a unit sphere and the volume V = 4 = π/3 of a unit sphere. Measuring the length of segments
on the circle leads to new concepts like angle or curvature. Because the circumference of the unit circle in the
plane is L = 2π, angle questions are tied to the number π, which Archimedes already approximated by fractions.
Also volumes were among the first quantities, Mathematicians wanted to measure and compute. A problem
on Moscow papyrus dating back to 1850 BC explains the general formula h(a2 + ab + b2 )/3 for a truncated
pyramid with base length a, roof length b and height h. Archimedes achieved to compute the volume of the
sphere: place a cone inside a cylinder. The complement of the cone inside the cylinder has on each height h
the area π − πh2 . The half sphere cut at height h is a disc of radius (1 − h2 ) which has area π(1 − h2 ) too.
Since the slices at each height have the same area, the volume must be the same. The complement of the cone
inside the cylinder has volume π − π/3 = 2π/3, half the volume of the sphere.
The first geometric playground was planimetry, the geometry in the flat two dimensional space. Highlights
are Pythagoras theorem, Thales theorem, Hippocrates theorem, and Pappus theorem. Discoveries
in planimetry have been made later on: an example is the Feuerbach 9 point theorem from the 19th century.
Ancient Greek Mathematics is closely related to history. It starts with Thales goes over Euclid’s era at 500
BC and ends with the threefold destruction of Alexandria 47 BC by the Romans, 392 by the Christians and
640 by the Muslims. Geometry was also a place, where the axiomatic method was brought to mathematics:
theorems are proved from a few statements which are called axioms like the 5 axioms of Euclid:
105
1. Any two distinct points A, B determines a line through A and B.

2. A line segment [A, B] can be extended to a straight line containing the segment.
3. A line segment [A, B] determines a circle containing B and center A.
4. All right angles are congruent.
5. If lines L, M intersect with a third so that inner angles add up to < π, then L, M intersect.
Euclid wondered whether the fifth postulate can be derived from the first four and called theorems derived
from the first four the ”absolute geometry”. Only much later, with Karl-Friedrich Gauss and Janos Bolyai
and Nicolai Lobachevsky in the 19’th century in hyperbolic space the 5’th axiom does not hold. Indeed,
geometry can be generalized to non-flat, or even much more abstract situations. Basic examples are geometry
on a sphere leading to spherical geometry or geometry on the Poincare disc, a hyperbolic space. Both
of these geometries are non-Euclidean. Riemannian geometry, which is essential for general relativity
theory generalizes both concepts to a great extent. An example is the geometry on an arbitrary surface. Cur-
vatures of such spaces can be computed by measuring length alone, which is how long light needs to go from
one point to the next.
An important moment in mathematics was the merge of geometry with algebra: this giant step is often
attributed to René Descartes. Together with algebra, the subject leads to algebraic geometry which can
be tackled with computers: here are some examples of geometries which are determined from the amount of
symmetry which is allowed:
Euclidean geometry Properties invariant under a group of rotations and translations

Affine geometry Properties invariant under a group of affine transformations
Projective geometry Properties invariant under a group of projective transformations
Spherical geometry Properties invariant under a group of rotations
Conformal geometry Properties invariant under angle preserving transformations
Hyperbolic geometry Properties invariant under a group of Möbius transformations
Here are four pictures about the 4 special points in a triangle and with which we will begin the lecture. We will
see why in each of these cases, the 3 lines intersect in a common point. It is a manifestation of a symmetry
present on the space of all triangles. size of the distance of intersection points is constant 0 if we move on the
space of all triangular shapes. It’s Geometry!
106
OLIVER KNILL
Lecture 4: Number Theory

Number theory studies the structure of integers like prime numbers and solutions to Diophantine equations.
Gauss called it the ”Queen of Mathematics”. Here are a few theorems and open problems.
An integer larger than 1 which is divisible by 1 and itself only is called a prime number. The number
257885161 − 1 is the largest known prime number. It has 17425170 digits. Euclid proved that there are infinitely
many primes: [Proof. Assume there are only finitely many primes p1 < p2 < · · · < pn . Then n = p1 p2 · · · pn + 1
is not divisible by any p1 , . . . , pn . Therefore, it is a prime or divisible by a prime larger than pn .] Primes become
more sparse as larger as they get. An important result is the prime number theorem which states that the n’th
prime number has approximately the size n log(n). For example the n = 1012 ’th prime is p(n) = 29996224275833
and n log(n) = 27631021115928.545... and p(n)/(n log(n)) = 1.0856... Many questions about prime numbers
are unsettled: Here are four problems: the third uses the notation (∆a)n = |an+1 − an | to get the absolute
difference. For example: ∆2 (1, 4, 9, 16, 25...) = ∆(3, 5, 7, 9, 11, ...) = (2, 2, 2, 2, ...). Progress on prime gaps has
been done in 2013: pn+1 − pn is smaller than 100’000’000 eventually (Yitang Zhang). pn+1 − pn is smaller than
600 eventually (Maynard). The largest known gap is 1476 which occurs after p = 1425172824437699411.
Landau there are infinitely many primes of the form n2 + 1.

Twin prime there are infinitely many primes p such that p + 2 is prime.
Goldbach every even integer n > 2 is a sum of two primes.
Gilbreath If pn enumerates the primes, then (∆k p)1 = 1 for all k > 0.
√ √
Andrica The prime gap estimate pn+1 − pn < 1 holds for all n.
If the sum of the proper divisors of a n is equal to n, then n is called a perfect number. For example,
6 is perfect as its proper divisors 1, 2, 3 sum up to 6. All currently known perfect numbers are even. The
question whether odd perfect numbers exist is probably the oldest open problem in mathematics and not
settled. Perfect numbers were familiar to Pythagoras and his followers already. Calendar coincidences like that
we have 6 work days and the moon needs ”perfect” 28 days to circle the earth could have helped to promote
the ”mystery” of perfect number. Euclid of Alexandria (300-275 BC) was the first to realize that if 2p − 1
is prime then k = 2p−1 (2p − 1) is a perfect number: [Proof: let σ(n) be the sum of all factors of n, including
n. Now σ(2n − 1)2n−1 ) = σ(2n − 1)σ(2n−1 ) = 2n (2n − 1) = 2 · 2n (2n − 1) shows σ(k) = 2k and verifies
that k is perfect.] Around 100 AD, Nicomachus of Gerasa (60-120) classified in his work ”Introduction to
Arithmetic” numbers on the concept of perfect numbers and lists four perfect numbers. Only much later it
became clear that Euclid got all the even perfect numbers: Euler showed that all even perfect numbers are of
the form (2n − 1)2n−1 , where 2n − 1 is prime. The factor 2n − 1 is called a Mersenne prime. [Proof: Assume
N = 2k m is perfect where m is odd and k > 0. Then 2k+1 m = 2N = σ(N ) = (2k+1 − 1)σ(m). This gives
σ(m) = 2k+1 m/(2k+1 − 1) = m(1 + 1/(2k+1 − 1)) = m + m/(2k+1 − 1). Because σ(m) and m are integers, also
m/(2k+1 − 1) is an integer. It must also be a factor of m. The only way that σ(m) can be the sum of only
two of its factors is that m is prime and so 2k+1 − 1 = m.] The first 39 known Mersenne primes are of
the form 2n − 1 with n = 2, 3, 5, 7, 13, 17, 19, 31, 61, 89, 107, 127, 521, 607, 1279, 2203, 2281, 3217, 4253,
4423, 9689, 9941, 11213, 19937, 21701, 23209, 44497, 86243, 110503, 132049, 216091, 756839, 859433, 1257787,
1398269, 2976221, 3021377, 6972593, 13466917. There are 11 more known from which one does not know the
rank of the corresponding Mersenne prime: n = 20996011, 24036583, 25964951, 30402457, 32582657, 37156667,
42643801,43112609,57885161, 74207281,77232917. The last was found in December 2017 only. It is unknown
whether there are infinitely many.
107
A polynomial equations for which all coefficients and variables are integers is called a Diophantine equation.
The first Diophantine equation studied already by Babylonians is x2 + y 2 = z 2 . A solution (x, y, z) of this
equation in positive integers is called a Pythagorean triple. For example, (3, 4, 5) is a Pythagorean triple.
Since 1600 BC, it is known that all solutions to this equation are of the form (x, y, z) = (2st, s2 − t2 , s2 + t2 ) or
(x, y, z) = (s2 − t2 , 2st, s2 + t2 ), where s, t are different integers. [Proof. Either x or y has to be even because
if both are odd, then the sum x2 + y 2 is even but not divisible by 4 but the right hand side is either odd or
divisible by 4. Move the even one, say x2 to the left and write x2 = z 2 − y 2 = (z − y)(z + y), then the right
hand side contains a factor 4 and is of the form 4s2 t2 . Therefore 2s2 = z − y, 2t2 = z + y. Solving for z, y gives
z = s2 + t2 , y = s2 − t2 , x = 2st.]
Analyzing Diophantine equations can be difficult. Only 10 years ago, one has established that the Fermat
equation xn + y n = z n has no solutions with xyz 6= 0 if n > 2. Here are some open problems for Diophantine
equations. Are there nontrivial solutions to the following Diophantine equations?
x6 + y 6 + z 6 + u6 + v 6 = w6 x, y, z, u, v, w > 0
x5 + y 5 + z 5 = w 5 x, y, z, w > 0
xk + y k = n!z k k ≥ 2, n > 1
xa + y b = z c , a, b, c > 2 gcd(a, b, c) = 1
The last equation is called Super Fermat. A Texan banker Andrew Beals once sponsored a prize of 1000 000
dollars for a proof or counter example to the statement: ”If xp + y q = z r with p, q, r > 2, then gcd(x, y, z) > 1.”
Given a prime like 7 and a number n we can add or subtract multiples of 7 from n to get a number in
{0, 1, 2, 3, 4, 5, 6 }. We write for example 19 = 12 mod 7 because 12 and 19 both leave the rest 5 when dividing
by 7. Or 5 ∗ 6 = 2 mod 7 because 30 leaves the rest 2 when dividing by 7. The most important theorem in
elementary number theory is Fermat’s little theorem which tells that if a is an integer and p is prime then
ap − a is divisible by p. For example 27 − 2 = 126 is divisible by 7. [Proof: use induction. For a = 0 it is clear.
The binomial expansion shows that (a+1)p −ap −1 is divisible by p. This means (a+1)p −(a+1) = (ap −a)+mp
for some m. By induction, ap − a is divisible by p and so (a + 1)p − (a + 1).] An other beautiful theorem is
Wilson’s theorem which allows to characterize primes: It tells that (n − 1)! + 1 is divisible by n if and only
if n is a prime number. For example, for n = 5, we verify that 4! + 1 = 25 is divisible by 5. [Proof: assume
n is prime. There are then exactly two numbers 1, −1 for which x2 − 1 is divisible by n. The other numbers
in 1, . . . , n − 1 can be paired as (a, b) with ab = 1. Rearranging the product shows (n − 1)! = −1 modulo n.
Conversely, if n is not prime, then n = km with k, m < n and (n − 1)! = ...km is divisible by n = km. ]
The solution to systems of linear equations like x = 3 (mod 5), x = 2 (mod 7) is given by the Chinese
remainder theorem. To solve it, continue adding 5 to 3 until we reach a number which leaves rest 2 to 7:
on the list 3, 8, 13, 18, 23, 28, 33, 38, the number 23 is the solution. Since 5 and 7 have no common divisor, the
system of linear equations has a solution.
For a given n, how do we solve x2 − yn = 1 for the unknowns y, x? A solution produces a square root x of 1
modulo n. For prime n, only x = 1, x = −1 are the solutions. For composite n = pq, more solutions x = r · s
where r2 = −1 mod p and s2 = −1 mod q appear. Finding x is equivalent to factor n, because the greatest
common divisor of x2 − 1 and n is a factor of n. Factoring is difficult if the numbers are large. It assures
that encryption algorithms work and that bank accounts and communications stay safe. Number theory,
once the least applied discipline of mathematics has become one of the most applied one in mathematics.
108
OLIVER KNILL
Lecture 5: Algebra
Algebra studies algebraic structures like ”groups” and ”rings”. The theory allows to solve polynomial
equations, characterize objects by its symmetries and is the heart and soul of many puzzles. Lagrange claims
Diophantus to be the inventor of Algebra, others argue that the subject started with solutions of quadratic
equation by Mohammed ben Musa Al-Khwarizmi in the book Al-jabr w’al muqabala of 830 AD. Solutions
to equation like x2 + 10x = 39 are solved there by completing the squares: add 25 on both sides go get
x2 + 10x + 25 = 64 and so (x + 5) = 8 so that x = 3.
The use of variables introduced in school in elementary algebra were introduced later. Ancient texts only
dealt with particular examples and calculations were done with concrete numbers in the realm of arithmetic.
Francois Viete (1540-1603) used first letters like A, B, C, X for variables.
The search for formulas for polynomial equations of degree 3 and 4 lasted 700 years. In the 16’th century,
the cubic equation and quartic equations were solved. Niccolo Tartaglia and Gerolamo Cardano reduced
the cubic to the quadratic: [first remove the quadratic part with X = x − a/3 so that X 3 + aX 2 + bX + c
becomes the depressed cubic x3 + px + q. Now substitute x = u − p/(3u) to get a quadratic equation
(u6 + qu3 − p3 /27)/u3 = 0 for u3 .] Lodovico Ferrari shows that the quartic equation can be reduced to the
cubic. For the quintic however no formulas could be found. It was Paolo Ruffini, Niels Abel and Évariste
Galois who independently realized that there are no formulas in terms of roots which allow to ”solve” equations
p(x) = 0 for polynomials p of degree larger than 4. This was an amazing achievement and the birth of ”group
theory”.
Two important algebraic structures are groups and rings.
In a group G one has an operation ∗, an inverse a−1 and a one-element 1 such that a ∗ (b ∗ c) = (a ∗ b) ∗ c, a ∗ 1 =
1 ∗ a = a, a ∗ a−1 = a−1 ∗ a = 1. For example, the set Q∗ of nonzero fractions p/q with multiplication operation ∗
and inverse 1/a form a group. The integers with addition and inverse a−1 = −a and ”1”-element 0 form a group
too. A ring R has two compositions + and ∗, where the plus operation is a group satisfying a+b = b+a in which
the one element is called 0. The multiplication operation ∗ has all group properties on R∗ except the existence
of an inverse. The two operations + and ∗ are glued together by the distributive law a ∗ (b + c) = a ∗ b + a ∗ c.
An example of a ring are the integers or the rational numbers or the real numbers. The later two are
actually fields, rings for which the multiplication on nonzero elements is a group too. The ring of integers are
no field because an integer like 5 has no multiplicative inverse. The ring of rational numbers however form a field.
Why is the theory of groups and rings not part of arithmetic? First of all, a crucial ingredient of algebra is
the appearance of variables and computations with these algebras without using concrete numbers. Second,
the algebraic structures are not restricted to ”numbers”. Groups and rings are general structures and extend
for example to objects like the set of all possible symmetries of a geometric object. The set of all similarity
operations on the plane for example form a group. An important example of a ring is the polynomial ring
of all polynomials. Given any ring R and a variable x, the set R[x] consists of all polynomials with coefficients
in R. The addition and multiplication is done like in (x2 + 3x + 1) + (x − 7) = x2 + 4x − 7. The problem to
factor a given polynomial with integer coefficients into polynomials of smaller degree: x2 − x + 2 for example
can be written as (x + 1)(x − 2) have a number theoretical flavor. Because symmetries of some structure form
a group, we also have intimate connections with geometry. But this is not the only connection with geometry.
Geometry also enters through the polynomial rings with several variables. Solutions to f (x, y) = 0 leads to
geometric objects with shape and symmetry which sometimes even have their own algebraic structure. They
are called varieties, a central object in algebraic geometry, objects which in turn have been generalized
109
further to schemes, algebraic spaces or stacks.
Arithmetic introduces addition and multiplication of numbers. Both form a group. The operations can be
written additively or multiplicatively. Lets look at this a bit closer: for integers, fractions and reals and the
addition +, the 1 element 0 and inverse −g, we have a group. Many groups are written multiplicatively where
the 1 element is 1. In the case of fractions or reals, 0 is not part of the multiplicative group because it is not
possible to divide by 0. The nonzero fractions or the nonzero reals form a group. In all these examples the
groups satisfy the commutative law g ∗ h = h ∗ g.
Here is a group which is not commutative: let G be the set of all rotations in space, which leave the unit
cube invariant. There are 3*3=9 rotations around each major coordinate axes, then 6 rotations around axes
connecting midpoints of opposite edges, then 2*4 rotations around diagonals. Together with the identity rotation
e, these are 24 rotations. The group operation is the composition of these transformations.
An other example of a group is S4 , the set of all permutations of four numbers (1, 2, 3, 4). If g : (1, 2, 3, 4) →
(2, 3, 4, 1) is a permutation and h : (1, 2, 3, 4) → (3, 1, 2, 4) is an other permutation, then we can combine the
two and define h ∗ g as the permutation which does first g and then h. We end up with the permutation
(1, 2, 3, 4) → (1, 2, 4, 3). The rotational symmetry group of the cube happens to be the same than the group
S4 . To see this ”isomorphism”, label the 4 space diagonals in the cube by 1, 2, 3, 4. Given a rotation, we can
look at the induced permutation of the diagonals and every rotation corresponds to exactly one permutation.
The symmetry group can be introduced for any geometric object. For shapes like the triangle, the cube, the
octahedron or tilings in the plane.
Symmetry groups describe geometric shapes by algebra.
Many puzzles are groups. A popular puzzle, the 15-puzzle was invented in 1874 by Noyes Palmer Chapman
in the state of New York. If the hole is given the number 0, then the task of the puzzle is to order a given
random start permutation of the 16 pieces. To do so, the user is allowed to transposes 0 with a neighboring
piece. Since every step changes the signature s of the permutation and changes the taxi-metric distance d of 0
to the end position by 1, only situations with even s + d can be reached. It was Sam Loyd who suggested to
start with an impossible solution and as an evil plot to offer 1000 dollars for a solution. The 15 puzzle group
has 16!/2 elements and the ”god number” is between 152 and 208. The Rubik cube is an other famous puzzle,
which is a group. Exactly 100 years after the invention of the 15 puzzle, the Rubik puzzle was introduced in
1974. Its still popular and the world record is to have it solved in 5.55 seconds. All Cubes 2x2x2 to 7x7x7 in a
row have been solved in a total time of 6 minutes. For the 3x3x3 cube, the God number is now known to be
20: one can always solve it in 20 or less moves.
Many puzzles are groups.
A small Rubik type game is the ”floppy”, which is a third of the Rubik and which has only 192 elements. An
other example is the Meffert’s great challenge. Probably the simplest example of a Rubik type puzzle is
the pyramorphix. It is a puzzle based on the tetrahedron. Its group has only 24 elements. It is the group
of all possible permutations of the 4 elements. It is the same group as the group of all reflection and rotation
symmetries of the cube in three dimensions and also is relevant when understanding the solutions to the quartic
equation discussed at the beginning. The circle is closed.
110
OLIVER KNILL
Lecture 6: Calculus
Calculus generalizes the process of taking differences and taking sums. Differences measure change, sums
explore how quantities accumulate. The procedure of taking differences has a limit called derivative. The
activity of taking sums leads to the integral. Sum and difference are dual to each other and related in an
intimate way. In this lecture, we look first at a simple set-up, where functions are evaluated on integers and
where we do not take any limits.
Several dozen thousand years ago, numbers were represented by units like 1, 1, 1, 1, 1, 1, . . . . The units were
carved into sticks or bones like the Ishango bone It took thousands of years until numbers were represented
with symbols like 0, 1, 2, 3, 4, . . . . Using the modern concept of function, we can say f (0) = 0, f (1) = 1, f (2) =
2, f (3) = 3 and mean that the function f assigns to an input like 1001 an output like f (1001) = 1001. Now
look at Df (n) = f (n + 1) − f (n), the difference. We see that Df (n) = 1 for all n. We can also formalize the
summation process. If g(n) = 1 is the constant 1 function, then then Sg(n) = g(0) + g(1) + · · · + g(n − 1) =
1 + 1 + · · · + 1 = n. We see that Df = g and Sg = f . If we start with f (n) = n and apply summation on
that function Then Sf (n) = f (0) + f (1) + f (2) + · · · + f (n − 1) leading to the values 0, 1, 3, 6, 10, 15, 21, . . . .
The new function g = Sf satisfies g(1) = 1, g(2) = 3, g(2) = 6, etc. The values are called the triangular
numbers. From g we can get back f by taking difference: Dg(n) = g(n + 1) − g(n) = f (n). For example
Dg(5) = g(6) − g(5) = 15 − 10 = 5 which indeed is f (5). Finding a formula for the sum Sf (n) is not so easy.
Can you do it? When Karl-Friedrich Gauss was a 9 year old school kid, his teacher, a Mr. Büttner gave him
the task to sum up the first 100 numbers 1 + 2 + · · · + 100. Gauss found the answer immediately by pairing
things up: to add up 1 + 2 + 3 + · · · + 100 he would write this as (1 + 100) + (2 + 99) + · · · + (50 + 51) leading
to 50 terms of 101 to get for n = 101 the value g(n) = n(n − 1)/2 = 5050. Taking differences again is easier
Dg(n) = n(n + 1)/2 − n(n − 1)/2 = n = f (n). If we add up he triangular numbers we compute h = Sg which
has the first values 0, 1, 4, 10, 20, 35, ..... These are the tetrahedral numbers because h(n) balls are needed
to build a tetrahedron of side length n. For example, h(4) = 20 golf balls are needed to build a tetrahedron of
side length 4. The formula which holds for h is h(n) = n(n − 1)(n − 2)/6 . Here is the fundamental theorem
of calculus, which is the core of calculus:
Df (n) = f (n) − f (0), DSf (n) = f (n) .
Proof.
n−1
X
SDf (n) = [f (k + 1) − f (k)] = f (n) − f (0) ,
k=0
n−1
X n−1
X
DSf (n) = [ f (k + 1) − f (k)] = f (n) .
k=0 k=0
Rx
The process of adding up numbers will lead to the integral 0
f (x) dx . The process of taking differences will
d
lead to the derivative dx f (x) .
The familiar notation is
Rx d d
Rx
0 dt
f (t) dt = f (x) − f (0), dx 0
f (t) dt = f (x)
If we define [n]0 = 1, [n]1 = n, [n]2 = n(n − 1)/2, [n]3 = n(n − 1)(n − 2)/6 then D[n] = [1], D[n]2 = 2[n], D[n]3 =
3[n]2 and in general
d n
dx [x] = n[x]n−1
The calculus you have just seen, contains the essence of single variable calculus. This core idea will become
more powerful and natural if we use it together with the concept of limit.
111
Problem: The Fibonnacci sequence 1, 1, 2, 3, 5, 8, 13, 21, . . . satisfies the rule f (x) = f (x − 1) + f (x − 2). For
example, f (6) = 8. What is the function g = Df , if we assume f (0) = 0? We take the difference between
successive numbers and get the sequence of numbers 0, 1, 1, 2, 3, 5, 8, ... which is the same sequence again. We
see that Df (x) = f (x − 1) .
If we take the same function f but now but now compute the function h(n) = Sf (n), we get the sequence
1, 2, 4, 7, 12, 20, 33, .... What sequence is that? Solution: Because Df (x) = f (x − 1) we have f (x) − f (0) =
SDf (x) = Sf (x − 1) so that Sf (x) = f (x + 1) − f (1). Summing the Fibonnacci sequence produces the
Fibonnacci sequence shifted to the left with f (2) = 1 is subtracted. It has been relatively easy to find the
sum, because we knew what the difference operation did. This example shows: we can study differences to
understand sums.
Problem: The function f (n) = 2n is called the exponential function. We have for example f (0) = 1, f (1) =
2, f (2) = 4, . . . . It leads to the sequence of numbers
n= 0 1 2 3 4 5 6 7 8 ...
f(n)= 1 2 4 8 16 32 64 128 256 . . .
We can verify that f satisfies the equation Df (x) = f (x) . because Df (x) = 2x+1 − 2x = (2 − 1)2x = 2x .
This is an important special case of the fact that
The derivative of the exponential function is the exponential function itself.
The function 2x is a special case of the exponential function when the Planck constant is equal to 1. We will see
that the relation will hold for any h > 0 and also in the limit h → 0, where it becomes the classical exponential
function ex which plays an important role in science.
Calculus has many applications: computing areas, volumes, solving differential equations. It even has applica-
tions in arithmetic. Here is an example for illustration. It is a proof that π is irrational The theorem is due
to Johann Heinrich Lambert (1728-1777): We show here the proof by Ivan Niven is given in a book of Niven-
Zuckerman-Montgomery. It originally appeared in 1947 (Ivan Niven, Bull.Amer.Math.Soc. 53 (1947),509). The
proof illustrates how calculus can help to get results in arithmetic.
Proof. Assume π = a/b with positive integers a and b. For any positive integer n define
f (x) = xn (a − bx)n /n! .
We have f (x) = f (π − x) and
0 ≤ f (x) ≤ π n an /n!(∗)
for 0 ≤ x ≤ π. For all 0 ≤ j ≤ n, the j-th derivative of f is zero at 0 and π and for n <= j, the j-th derivative
of f is an integer at 0 and π.
The function F (x) = f (x) − f (2) (x) + f (4) (x) − ... + (−1)n f (2n) (x) has the property that F (0) and F (π) are
integers and F + F 00 = f . Therefore, (F 0 (x) sin(x) − F (x) cos(x))0 = f sin(x). By the fundamental theorem of
Rπ
calculus, 0 f (x) sin(x) dx is an integer. Inequality (*) implies however that this integral is between 0 and 1 for
large enough n. For such an n we get a contradiction.
112
OLIVER KNILL
Lecture 7: Set Theory and Logic

Set theory studies sets, the fundamental building blocks of mathematics. While logic describes the language of
all mathematics, set theory provides the framework for additional structures like category theory. In Cantorian
set theory, one can compute with subsets of a given set X like with numbers. There are two basic operations:
the addition A + B of two sets is defined as the set of all points which are in exactly one of the sets. The
multiplication A · B of two sets contains all the points which are in both sets. With the symmetric difference
as addition and the intersection as multiplication, the subsets of a given set X become a ring. This Boolean
ring has the property A + A = 0 and A · A = A for all sets. The zero element is the empty set ∅ = {}. The
additive inverse of A is the complement −A of A in X. The multiplicative 1-element is the set X because
X · A = A. As in the ring Z of integers, the addition and multiplication on sets is commutative. Multiplication
does not have an inverse in general. Two sets A, B have the same cardinality, if there exists a one-to-one map
from A to B. For finite sets, this means that they have the same number of elements. Sets which do not have
finitely many elements are called infinite. Do all sets with infinitely many elements have the same cardinality?
The integers Z and the natural numbers N for example are infinite sets which have the same cardinality: the
map f (2n) = n, f (2n + 1) = −n establishes a bijection between N and Z. Also the rational numbers Q have
the same cardinality than N. Associate a fraction p/q with a point (p, q) in the plane. Now cut out the column
q = 0 and run the Ulam spiral on the modified plane. This provides a numbering of the rationals. Sets which
can be counted are called of cardinality ℵ0 . Does an interval have the same cardinality than the reals? Even
so an interval like I = (−π/2, π/2) has finite length, one can bijectively map it to R with the tan function as
tan : I → R is bijective. Similarly, one can see that any two intervals of positive length have the same cardinality.
It was a great moment of mathematics, when Georg Cantor realized in 1874 that the interval (0, 1) does not
have the same cardinality than the natural numbers. His argument is ingenious: assume, we could count the
points a1 , a2 , . . . . If 0.ai1 ai2 ai3 ... is the decimal expansion of ai , define the real number b = 0.b1 b2 b3 ..., where
bi = aii + 1 mod 10. Because this number b does not agree at the first decimal place with a1 , at the second
place with a2 and so on, the number b does not appear in that enumeration of all reals. It has positive distance
at least 10−i from the i’th number (and any representation of the number by a decimal expansion which is
equivalent). This is a contradiction. The new cardinality, the continuum is also denoted ℵ1 . The reals are
uncountable. This gives elegant proofs like the existence of transcendental number, numbers which are not
algebraic, meaning that they are not the root of any polynomial with integer coefficients: algebraic numbers can
be counted. Similarly as one can establish a bijection between the natural numbers N and the integers Z, there
is a bijection f between the interval I and the unit square: if x = 0.x1 x2 x3 . . . is the decimal expansion of x then
f (x) = (0.x1 x3 x5 . . . , 0.x2 x4 x6 . . . ) is the bijection. Are there cardinalities larger than ℵ1 ? Cantor answered
also this question. He showed that for an infinite set, the set of all subsets has a larger cardinality than the set
itself. How does one see this? Assume there is a bijection x → A(x) which maps each point to a set A(x). Now
look at the set B = {x | x ∈ / A(x) } and let b be the point in X which corresponds to B. If y ∈ B, then y ∈ / B(x).
On the other hand, if y ∈ / B, then y ∈ B. The set B does appear in the ”enumeration” x → A(x) of all sets. The
set of all subsets of N has the same cardinality than the continuum: A → j∈A 1/2j provides a map from P (N )
P
to [0, 1]. The set of all finite subsets of N however can be counted. The set of all subsets of the real numbers
has cardinality ℵ2 , etc. Is there a cardinality between ℵ0 and ℵ1 ? In other words, is there a set which can
not be counted and which is strictly smaller than the continuum in the sense that one can not find a bijection
between it and R? This was the first of the 23 problems posed by Hilbert in 1900. The answer is surprising:
one has a choice. One can accept either the ”yes” or the ”no” as a new axiom. In both cases, Mathematics
is still fine. The nonexistence of a cardinality between ℵ0 and ℵ1 is called the continuum hypothesis and
is usually abbreviated CH. It is independent of the other axioms making up mathematics. This was the work
113
of Kurt Gödel in 1940 and Paul Cohen in 1963. The story of exploring the consistency and completeness
of axiom systems of all of mathematics is exciting. Euclid axiomatized geometry, Hilbert’s program was more
ambitious. He aimed at a set of axiom systems for all of mathematics. The challenge to prove Euclid’s 5’th
postulate is paralleled by the quest to prove the CH. But the later is much more fundamental because it deals
with all of mathematics and not only with some geometric space. Here are the Zermelo-Frenkel Axioms
(ZFC) including the Axiom of choice (C) as established by Ernst Zermelo in 1908 and Adolf Fraenkel and
Thoral Skolem in 1922.
Extension If two sets have the same elements, they are the same.
Image Given a function and a set, then the image of the function is a set too.
Pairing For any two sets, there exists a set which contains both sets.
Property For any property, there exists a set for which each element has the property.
Union Given a set of sets, there exists a set which is the union of these sets.
Power Given a set, there exists the set of all subsets of this set.
Infinity There exists an infinite set.
Regularity Every nonempty set has an element which has no intersection with the set.
Choice Any set of nonempty sets leads to a set which contains an element from each.
There are other systems like ETCS, which is the elementary theory of the category of sets. In category
theory, not the sets but the categories are the building blocks. Categories do not form a set in general. It
elegantly avoids the Russel paradox too. The axiom of choice (C) has a nonconstructive nature which can
lead to seemingly paradoxical results like the Banach Tarski paradox: one can cut the unit ball into 5 pieces,
rotate and translate the pieces to assemble two identical balls of the same size than the original ball. Gödel and
Cohen showed that the axiom of choice is logically independent of the other axioms ZF. Other axioms in ZF
have been shown to be independent, like the axiom of infinity. A finitist would refute this axiom and work
without it. It is surprising what one can do with finite sets. The axiom of regularity excludes Russellian
sets like the set X of all sets which do not contain themselves. The Russell paradox is: Does X contain
X? It is popularized as the Barber riddle: a barber in a town only shaves the people who do not shave
themselves. Does the barber shave himself? Gödels theorems of 1931 deal with mathematical theories
which are strong enough to do basic arithmetic in them.
First incompleteness theorem: Second incompleteness theorem:
In any theory there are true statements which can In any theory, the consistency of the theory can not
not be proved within the theory. be proven within the theory.
The proof uses an encoding of mathematical sentences which allows to state liar paradoxical statement ”this
sentence can not be proved”. While the later is an odd recreational entertainment gag, it is the core for a theorem
which makes striking statements about mathematics. These theorems are not limitations of mathematics; they
illustrate its infiniteness. How awful if one could build axiom system and enumerate mechanically all possible
truths from it.
114
OLIVER KNILL
Lecture 8: Probability theory

Probability theory is the science of chance. It starts with combinatorics and leads to a theory of stochas-
tic processes. Historically, probability theory initiated from gambling problems as in Girolamo Cardano’s
gamblers manual in the 16th century. A great moment of mathematics occurred, when Blaise Pascal and
Pierre Fermat jointly laid a foundation of mathematical probability theory.
It took a while to formalize “randomness” precisely. Here is the setup as which it had been put forward by
Andrey Kolmogorov: all possible experiments of a situation are modeled by a set Ω, the ”laboratory”. A
measurable subset of experiments is called an “event”. Measurements are done by real-valued functions X.
These functions are called random variables and are used to observe the laboratory.
As an example, let us model the process of throwing a coin 5 times. An experiment is a word like httht, where h
stands for “head” and t represents “tail”. The laboratory consists of all such 32 words. We could look for example
at the event A that the first two coin tosses are tail. It is the set A = {ttttt, tttth, tttht, ttthh, tthtt, tthth, tthht, tthhh}.
We could look at the random variable which assigns to a word the number of heads. For every experiment, we
get a value, like for example, X[tthht] = 2.
In order to make statements about randomness, the concept of a probability measure is needed. This is
a function P from the set of all events to the interval [0, 1]. It should have the property that P [Ω] = 1 and
P [A1 ∪ A2 ∪ · · · ] = P [A1 ] + P [A2 ] + · · · , if Ai is a sequence of disjoint events.
The most natural probability measure on a finite set Ω is P [A] = kAk/kΩk, where kAk stands for the number
of elements in A. It is the “number of good cases” divided by the “number of all cases”. For example, to count
the probability of the event A that we throw 3 heads during the 5 coin tosses, we have |A| = 10 possibilities.
Since the entire laboratory has |Ω| = 32 possibilities, the probability of the event is 10/32. In order to study
these probabilities, one needs combinatorics:
How many ways are there to: The answer is:

rearrange or permute n elements n! = n(n − 1)...2 · 1
choose k from n with repetitions nk
n!
pick k from n if order matters (n−k)!!
n n!
pick k from n with order irrelevant = k!(n−k)!
k
P
The expectation of a random variable E[X] is defined as the sum m = ω∈Ω X(ω)P [{ω}]. In our coin toss
experiment, this is 5/2. The variance of X is the expectation of (X − m)2 . In our coin experiments, it is 5/4.
The square root of the variance is the standard deviation. This is the expected deviation from the mean. An
event happens almost surely if the event has probability 1.
2
An important case of a random variable is X(ω) = ω on Ω = R equipped with probability P [A] = A √1π e−x dx,
R
the standard normal distribution. Analyzed first by Abraham de Moivre in 1733, it was studied by Carl
Friedrich Gauss in 1807 and therefore also called Gaussian distribution.
Two random variables X, Y are called uncorrelated, if E[XY ] = E[X] · E[Y ]. If for any functions f, g also
f (X) and g(Y ) are uncorrelated, then X, Y are called independent. Two random variables are said to have
the same distribution, if for any a < b, the events {a ≤ X ≤ b } and {a ≤ Y ≤ b } are independent. If X, Y
are uncorrelated, then the relation Var[X] + Var[Y ] = Var[X + Y ] holds which is just Pythagoras theorem,
because uncorrelated can be understood geometrically: X − E[X] and Y − E[Y ] are orthogonal. A common
problem is to study the sum of independent random variables Xn with identical distribution. One abbreviates
this IID. Here are the three most important theorems which we formulate in the case, where all random variables
are assumed to have expectatation 0 and standard deviation 1. Let Sn = X1 + ... + Xn be the n’th sum of the
115
IID random variables. It is also called a random walk.
LLN Law of Large Numbers assures that Sn /n converges to 0.

√
CLT Central Limit Theorem:Sn / n approaches the Gaussian distribution.
p
LIL Law of Iterated Logarithm: Sn / 2n log log(n) accumulates in [−1, 1].
The LLN shows that one can find out about the expectation by averaging experiments. The CLT explains why
one sees the standard normal distribution so often. The LIL finally gives us a precise estimate how fast Sn
grows. Things become interesting if the random variables are no more independent. Generalizing LLN,CLT,LIL
to such situations is part of ongoing research.
Here are two open questions in probability theory:
√
Are numbers like π, e, 2 normal: do all digits appear with the same frequency?
What growth rates Λn can occur in Sn /Λn having limsup 1 and liminf −1?
p
For the second question, there are examples for Λn = 1, λn = log(n) and of course λn = n log log(n) from
LIL if the random variables are independent. Examples of random variables which are not independent are
√
Xn = cos(n 2).
Statistics is the science of modeling random events in a probabilistic setup. Given data points, we want to
find a model which fits the data best. This allows to understand the past, predict the future or discover
laws of nature. The most common task is to find the mean and the standard deviation of some data. The
Pn Pn
mean is also called the average and given by m = n1 k=1 xk . The variance is σ 2 = n1 k=1 (xk − m)2 with
standard deviation σ.
A sequence of random variables Xn define a so called stochastic process. Continuous versions of such pro-
cesses are where Xt is a curve of random random variables. An important example is Brownian motion,
which is a model of a random particles.
Besides gambling and analyzing data, also physics was an important motivator to develop probability theory.
An example is statistical mechanics, where the laws of nature are studied with probabilistic methods. A
famous physical law is Ludwig Boltzmann’s relation S = k log(W ) for entropy, a formula which decorates
Boltzmann’s tombstone. The entropy of a probability measure P [{k}] = pk on a finite set {1, ..., n} is defined
Pn
as S = − i=1 pi log(pi ). Today, we would reformulate Boltzmann’s law and say that it is the expectation
S = E[log(W )] of the logarithm of the “Wahrscheinlichkeit” random variable W (i) = 1/pi on Ω = {1, ..., n }.
Entropy is important because nature tries to maximize it
116
OLIVER KNILL
Lecture 9: Topology
Topology studies properties of geometric objects which do not change under continuous reversible deforma-
tions. In topology, a coffee cup with a single handle is the same as a doughnut. One can deform one into the
other without punching any holes in it or ripping it apart. Similarly, a plate and a croissant are the same. But
a croissant is not equivalent to a doughnut. On a doughnut, there are closed curves which can not be pulled
together to a point. For a topologist the letters O and P are the equivalent but different from the letter B.
The mathematical setup is beautiful: a topological space is a set X with a set O of subsets of X containing
both ∅ and X such that finite intersections and arbitrary unions in O are in O. Sets in O are called open sets
and O is called a topology. The complement of an open set is called closed. Examples of topologies are the
trivial topology O = {∅, X}, where no open sets besides the empty set and X exist or the discrete topology
O = {A | A ⊂ X}, where every subset is open. But these are in general not interesting. An important example
on the plane X is the collection O of sets U in the plane X for which every point is the center of a small disc
still contained in U . A special class of topological spaces are metric spaces, where a set X is equipped with a
distance function d(x, y) = d(y, x) ≥ 0 which satisfies the triangle inequality d(x, y) + d(y, z) ≥ d(x, z) and
for which d(x, y) = 0 if and only if x = y. A set U in a metric space is open if to every x in U , there is a ball
Br (x) = {y|d(x, y) < r} of positive radius r contained in U . Metric spaces are topological spaces but not vice
versa: the trivial topology for example is not in general. For doing calculus on a topological space X, each
point has a neighborhood called chart which is topologically equivalent to a disc in Euclidean space. Finitely
many neighborhoods covering X form an atlas of X. If the charts are glued together with identification maps
on the intersection one obtains a manifold. Two dimensional examples are the sphere, the torus, the pro-
jective plane or the Klein bottle. Topological spaces X, Y are called homeomorphic meaning “topologically
equivalent” if there is an invertible map from X to Y such that this map induces an invertible map on the
corresponding topologies. How can one decide whether two spaces are equivalent in this sense? The surface of
the coffee cup for example is equivalent in this sense to the surface of a doughnut but it is not equivalent to the
surface of a sphere. Many properties of geometric spaces can be understood by discretizing it like with a graph.
A graph is a finite collection of vertices V together with a finite set of edges E, where each edge connects two
points in V . For example, the set V of cities in the US where the edges are pairs of cities connected by a street
is a graph. The Königsberg bridge problem was a trigger puzzle for the study of graph theory. Polyhedra
were an other start in graph theory. It study is loosely related to the analysis of surfaces. The reason is that
one can see polyhedra as discrete versions of surfaces. In computer graphics for example, surfaces are rendered
as finite graphs, using triangularizations. The Euler characteristic of a convex polyhedron is a remarkable
topological invariant. It is V − E + F = 2, where V is the number of vertices, E the number of edges and F the
number of faces. This number is equal to 2 for connected polyhedra in which every closed loop can be pulled
together to a point. This formula for the Euler characteristic is also called Euler’s gem. It comes with a rich
history. René Descartes stumbled upon it and written it down in a secret notebook. It was Leonard Euler
in 1752 was the first to proved the formula for convex polyhedra. A convex polyhedron is called a Platonic
solid, if all vertices are on the unit sphere, all edges have the same length and all faces are congruent polygons.
A theorem of Theaetetus states that there are only five Platonic solids: [Proof: Assume the faces are regular
n-gons and m of them meet at each vertex. Beside the Euler relation V + E + F = 2, a polyhedron also satisfies
the relations nF = 2E and mV = 2E which come from counting vertices or edges in different ways. This gives
2E/m − E + 2E/n = 2 or 1/n + 1/m = 1/E + 1/2. From n ≥ 3 and m ≥ 3 we see that it is impossible that both
m and n are larger than 3. There are now nly two possibilities: either n = 3 or m = 3. In the case n = 3 we
have m = 3, 4, 5 in the case m = 3 we have n = 3, 4, 5. The five possibilities (3, 3), (3, 4), (3, 5), (4, 3), (5, 3)
117
represent the five Platonic solids.] The pairs (n, m) are called the Schläfly symbol of the polyhedron:
Name V E F V-E+F Schläfli Name V E F V-E+F Schläfli
tetrahedron 4 6 4 2 {3, 3}
hexahedron 8 12 6 2 {4, 3} dodecahedron 20 30 12 2 {5, 3}
octahedron 6 12 8 2 {3, 4} icosahedron 12 30 20 2 {3, 5}
The Greeks proceeded geometrically: Euclid showed in the ”Elements” that each vertex can have either 3,4 or 5
equilateral triangles attached, 3 squares or 3 regular pentagons. (6 triangles, 4 squares or 4 pentagons would lead
to a total angle which is too large because each corner must have at least 3 different edges). Simon Antoine-
Jean L’Huilier refined in 1813 Euler’s formula to situations with holes: V − E + F = 2 − 2g ,
where g is the number of holes. For a doughnut it is V − E + F = 0. Cauchy first proved that there are 4
non-convex regular Kepler-Poinsot polyhedra.
Name V E F V-E+F Schläfli

small stellated dodecahedron 12 30 12 -6 {5/2, 5}
great dodecahedron 12 30 12 -6 {5, 5/2}
great stellated dodecahedron 20 30 12 2 {5/2, 3}
great icosahedron 12 30 20 2 {3, 5/2}
If two different face types are allowed but each vertex still look the same, one obtains 13 semi-regular polyhe-
dra. They were first studied by Archimedes in 287 BC. Since his work is lost, Johannes Kepler is considered
the first since antiquity to describe all of them them in his ”Harmonices Mundi”. The Euler characteristic for
surfaces is χ = 2−2g where g is the number of holes. The computation can be done by triangulating the surface.
The Euler characteristic characterizes smooth compact surfaces if they are orientable. A non-orientable surface,
the Klein bottle can be obtained by gluing ends of the Möbius strip. Classifying higher dimensional manifolds
is more difficult and finding good invariants is part of modern research. Higher analogues of polyhedra are
called polytopes (Alicia Boole Stott). Regular polytopes are the analogue of the Platonic solids in higher
dimensions. Examples:
dimension name Schläfli symbols

2: Regular polygons {3}, {4}, {5}, ...
3: Platonic solids {3, 3}, {3, 4}, {3, 5}, {4, 3}, {5, 3}
4: Regular 4D polytopes {3, 3, 3}, {4, 3, 3}, {3, 3, 4}, {3, 4, 3}, {5, 3, 3}, {3, 3, 5}
≥ 5: Regular polytopes {3, 3, 3, . . . , 3}, {4, 3, 3, . . . , 3}, {3, 3, 3, . . . , 3, 4}
Ludwig Schll̈afly saw in 1852 exactly six convex regular convex 4-polytopes or polychora, where ”Choros”
is Greek for ”space”. Schlaefli’s polyhedral formula is V − E + F − C = 0 holds, where C
is the number of 3-dimensional chambers. In dimensions 5 and higher, there are only 3 types of poly-
topes: the higher dimensional analogues of the tetrahedron, octahedron and the cube. A general formula
Pd−1 k d
k=0 (−1) vk = 1 − (−1)
gives the Euler characteristic of a convex polytop in d dimensions with
k-dimensional parts vk .
118
OLIVER KNILL
Lecture 10: Analysis
Analysis is a science of measure and optimization. As a rather diverse collection of mathematical fields, it
contains real and complex analysis, functional analysis, harmonic analysis and calculus of variations.
Analysis has relations to calculus, geometry, topology, probability theory and dynamical systems. We focus
here mostly on ”the geometry of fractals” which can be seen as part of dimension theory. Examples are Julia
sets which belong to the subfield of ”complex analysis” of ”dynamical systems”. ”Calculus of variations” is
illustrated by the Kakeya needle set in ”geometric measure theory”, ”Fourier analysis” appears when looking
at functions which have fractal graphs, ”spectral theory” as part of functional analysis is represented by the
”Hofstadter butterfly”. We somehow describe the topic using ”pop icons”.
A fractal is a set with non-integer dimension. An example is the Cantor set, as discovered in 1875 by Henry
Smith. Start with the unit interval. Cut the middle third, then cut the middle third from both parts then the
middle parts of the four parts etc. The limiting set is the Cantor set. The mathematical theory of fractals belongs
to measure theory and can also be thought of a playground for real analysis or topology. The term fractal
had been introduced by Benoit Mandelbrot in 1975. Dimension can be defined in different ways. The simplest
is the box counting definition which works for most household fractals: if we need n squares of length r to
cover a set, then d = − log(n)/ log(r) converges to the dimension of the set with r → 0. A curve
of length L for example needs L/r squares of length r so that its dimension is 1. A region of area A needs A/r2
squares of length r to be covered and its dimension is 2. The Cantor set needs to be covered with n = 2m squares
of length r = 1/3m . Its dimension is − log(n)/ log(r) = −m log(2)/(m log(1/3)) = log(2)/ log(3). Examples of
fractals are the graph of the Weierstrass function 1872, the Koch snowflak (1904), the Sierpinski carpet (1915)
or the Menger sponge (1926).
Complex analysis extends calculus to the complex. It deals with functions f (z) defined in the complex plane.
Integration is done along paths. Complex analysis completes the understanding about functions. It also provides
more examples of fractals by iterating functions like the quadratic map f (z) = z 2 + c:
One has already iterated functions before like the Newton method (1879). The Julia sets were introduced in
1918, the Mandelbrot set in 1978 and the Mandelbar set in 1989. Particularly famous are the Douady rabbit
and the dragon, the dendrite, the airplane. Calculus of variations is calculus in infinite dimensions.
Taking derivatives is called taking ”variations”. Historically, it started with the problem to find the curve of
fastest fall leading to the Brachistochrone curve ~r(t) = (t − sin(t), 1 − cos(t)). In calculus, we find maxima
and minima of functions. In calculus of variations, we extremize on much larger spaces. Here are examples of
problems:
Brachistochrone 1696
Minimal surface 1760
Geodesics 1830
Isoperimetric problem 1838
Kakeya Needle problem 1917
Fourier theory decomposes a function into basic components of various frequencies f (x) = a1 sin(x) +
a2 sin(2x) + a3 sin(3x) + · · · . The numbers ai are called the Fourier coefficients. Our ear does such a
decomposition, when we listen to music. By distinguish different frequencies, our ear produces a Fourier anal-
ysis.
119
Fourier series 1729

Fourier transform (FT) 1811
Discrete FT Gauss?
Wavelet transform 1930
The Weierstrass function mentioned above is given as a series n an cos(πbn x) with 0 < a < 1, ab > 1 + 3π/2.
P
The dimension of its graph is believed to be 2 + log(a)/ log(b) but no rigorous computation of the dimension
was done yet. Spectral theory analyzes linear maps L. The spectrum are the real numbers E such that
L − E is not invertible. A Hollywood celebrity among all linear maps is the almost Matthieu operator
L(x)n = xn+1 + xn−1 + (2 − 2 cos(cn))xn : if we draw the spectrum for for each c, we see the Hofstadter
butterfly. For fixed c the map describes the behavior of an electron in an almost periodic crystal. An
other famous system is the quantum harmonic oscillator, L(f ) = f 00 (x) + f (x), the vibrating drum
L(f ) = fxx + fyy , where f is the amplitude of the drum and f = 0 on the boundary of the drum.
Hydrogen atom 1914

Hofstadter butterfly 1976
Harmonic oscillator 1900
Vibrating drum 1680
All these examples in analysis look unrelated at first. Fractal geometry ties many of them together: spectra are
often fractals, minimal configurations have fractal nature, like in solid state physics or in diffusion limited
aggregation or in other critical phenomena like percolation phenomena, cracks in solids or the formation
of lighting bolts In Hamiltonian mechanics, minimal energy configurations are often fractals like Mather
theory. And solutions to minimizing problems lead to fractals in a natural way like when you have the task to
turn around a needle on a table by 180 degrees and minimize the area swept out by the needle. The minimal
turn leads to a Kakaya set, which is a fractal. Finally, lets mention some unsolved problems in analysis: does the
P∞
Riemann zeta function f (z) = n=1 1/nz have all nontrivial roots on the axis Re(z) = 1/2? This question
is called the Riemann hypothesis and is the most important open problem in mathematics. It is an example
of a question in analytic number theory which also illustrates how analysis has entered into number theory.
Some mathematicians think that spectral theory might solve it. Also the Mandelbrot set M is not understood
yet: the ”holy grail” in the field of complex dynamics is the problem whether it M is locally connected. From
the Hofstadter butterfly one knows that it has measure zero. What is its dimension? An other open question
in spectral theory is the ”can one hear the sound of a drum” problem which asks whether there are two convex
drums which are not congruent but which have the same spectrum. In the area of calculus of variations, just one
problem: how long is the shortest curve in space such that its convex hull (the union of all possible connections
between two points on the curve) contains the unit ball.
120
OLIVER KNILL
Lecture 11: Cryptography

Cryptography is the theory of codes. Two important aspects of the field are the encryption rsp. decryption
of information and error correction. Both are crucial in daily life. When getting access to a computer, viewing
a bank statement or when taking money from the ATM, encryption algorithms are used. When phoning, surfing
the web, accessing data on a computer or listening to music, error correction algorithms are used. Since our
lives have become more and more digital: music, movies, books, journals, finance, transportation, medicine,
and communication have become digital, we rely on strong error correction to avoid errors and encryption
to assure things can not be tempered with. Without error correction, airplanes would crash: small errors
in the memory of a computer would produce glitches in the navigation and control program. In a computer
memory every hour a couple of bits are altered, for example by cosmic rays. Error correction assures that this
gets fixed. Without error correction music would sound like a 1920 gramophone record. Without encryption,
everybody could intrude electronic banks and transfer money. Medical history shared with your doctor would
all be public. Before the digital age, error correction was assured by extremely redundant information storage.
Writing a letter on a piece of paper displaces billions of billions of molecules in ink. Now, changing any single
bit could give a letter a different meaning. Before the digital age, information was kept in well guarded safes
which were physically difficult to penetrate. Now, information is locked up in computers which are connected
to other computers. Vaults, money or voting ballots are secured by mathematical algorithms which assure
that information can only be accessed by authorized users. Also life needs error correction: information in the
genome is stored in a genetic code, where a error correction makes sure that life can survive. A cosmic ray
hitting the skin changes the DNA of a cell, but in general this is harmless. Only a larger amount of radiation
can render cells cancerous.
How can an encryption algorithm be safe? One possibility is to invent a new method and keep it secret. An
other is to use a well known encryption method and rely on the difficulty of mathematical computation
tasks to assure that the method is safe. History has shown that the first method is unreliable. Systems which
rely on ”security through obfuscation” usually do not last. The reason is that it is tough to keep a method
secret if the encryption tool is distributed. Reverse engineering of the method is often possible, for example
using plain text attacks. Given a map T , a third party can compute pairs x, T (x) and by choosing specific texts
figure out what happens.
The Caesar cypher permutes the letters of the alphabet. We can for example replace every letter A with
B, every letter B with C and so on until finally Z is replaced with A. The word ”Mathematics” becomes so
encrypted as ”Nbuifnbujdt”. Caesar would shift the letters by 3. The right shift just discussed was used by
his Nephew Augustus. Rot13 shifts by 13, and Atbash cypher reflects the alphabet, switch A with Z, B
with Y etc. The last two examples are involutive: encryption is decryption. More general cyphers are obtained
by permuting the alphabet. Because of 26! = 403291461126605635584000000 ∼ 1027 permutations, it appears
first that a brute force attack is not possible. But Cesar cyphers can be cracked very quickly using statistical
analysis. If we know the frequency with which letters appear and match the frequency of a text we can figure
out which letter was replaced with which. The Trithemius cypher prevents this simple analysis by changing
the permutation in each step. It is called a polyalphabetic substitution cypher. Instead of a simple permutation,
there are many permutations. After transcoding a letter, we also change the key. Lets take a simple example.
Rotate for the first letter the alphabet by 1, for the second letter, the alphabet by 2, for the third letter, the
alphabet by 3 etc. The word ”Mathematics” becomes now ”Ncwljshbrmd”. Note that the second ”a” has been
translated to something different than a. A frequency analysis is now more difficult. The Viginaire cypher
adds even more complexity: instead of shifting the alphabet by 1, we can take a key like ”BCNZ”, then shift the
first letter by 1, the second letter by 3 the third letter by 13, the fourth letter by 25 the shift the 5th letter by
121
1 again. While this cypher remained unbroken for long, a more sophisticated frequency analysis which involves
first finding the length of the key makes the cypher breakable. With the emergence of computers, even more
sophisticated versions like the German enigma had no chance.
Diffie-Hellman key exchange allows Ana and Bob want to agree on a secret key over a public channel. The
two palindromic friends agree on a prime number p and a base a. This information can be exchanged over an
open channel. Ana chooses now a secret number x and sends X = ax modulo p to Bob over the channel. Bob
chooses a secret number y and sends Y = ay modulo p to Ana. Ana can compute Y x and Bob can compute
X y but both are equal to axy . This number is their common secret. The key point is that eves dropper Eve,
can not compute this number. The only information available to Eve are X and Y , as well as the base a and p.
Eve knows that X = ax but can not determine x. The key difficulty in this code is the discrete log problem:
getting x from ax modulo p is believed to be difficult for large p.
The Rivest-Shamir-Adleman public key system uses a RSA public key (n, a) with an integer n = pq and
a < (p − 1)(q − 1), where p, q are prime. Also here, n and a are public. Only the factorization of n is kept secret.
Ana publishes this pair. Bob who wants to email Ana a message x, sends her y = xa mod n. Ana, who has
computed b with ab = 1 mod (p−1)(q−1) can read the secrete email y because y b = xab = x(p−1)(q−1) = x modn.
But Eve, has no chance because the only thing Eve knows is y and (n, a). It is believed that without the
factorization of n, it is not possible to determine x. The message has been transmitted securely. The core
difficulty is that taking roots in the ring Zn = {0, . . . , n − 1 } is difficult without knowing the factorization
of n. With a factorization, we can quickly take arbitrary roots. If we can take square roots, then we can also
factor: assume we have a product n = pq and we know how to take square roots of 1. If x solves x2 = 1 mod n
and x is different from 1, then x2 − 1 = (x − 1)(x + 1) is zero modulo n. This means that p divides (x − 1) or
(x + 1). To find a factor, we can take the greatest common divisor of n, x − 1. Take n = 77 for example. We
are given the root 34 of 1. ( 342 = 1156 has reminder 1 when divided by 34). The greatest common divisor of
34 − 1 and 77 is 11 is a factor of 77. Similarly, the greatest common divisor of 34 + 1 and 77 is 7 divides 77.
Finding roots modulo a composite number and factoring the number is equally difficult.
Cipher Used for Difficulty Attack
Cesar transmitting messages many permutations Statistics
Viginere transmitting messages many permutations Statistics
Enigma transmitting messages no frequency analysis Plain text
Diffie-Helleman agreeing on secret key discrete log mod p Unsafe primes
RSA electronic commerce factoring integers Factoring
The simplest error correcting code uses 3 copies of the same information so single error can be corrected.
With 3 watches for example, one watch can fail. But this basic error correcting code is not efficient. It can
correct single errors by tripling the size. Its efficiency is 33 percent.
122
OLIVER KNILL
Lecture 12: Dynamical systems

Dynamical systems theory is the science of time evolution. If time is continuous the evolution is defined
by a differential equation ẋ = f (x). If time is discrete then we look at the iteration of a map x → T (x).
The goal of the theory is to predict the future of the system when the present state is known. A differential
equation is an equation of the form d/dtx(t) = f (x(t)), where the unknown quantity is a path x(t) in some
“phase space”. We know the velocity d/dtx(t) = ẋ(t) at all times and the initial configuration x(0)), we can to
compute the trajectory x(t). What happens at a future time? Does x(t) stay in a bounded region or escape
to infinity? Which areas of the phase space are visited and how often? Can we reach a certain part of the
space when starting at a given point and if yes, when. An example of such a question is to predict, whether an
asteroid located at a specific location will hit the earth or not. An other example is to predict the weather of
the next week.
An examples of a dynamical systems in one dimension is the differential equation
x0 (t) = x(t)(2 − x(t)), x(0) = 1
It is called the logistic system and describes population growth. This system has the solution x(t) =
2et /(1 + e2t ) as you can see by computing the left and right hand side.
A map is a rule which assigns to a quantity x(t) a new quantity x(t + 1) = T (x(t)). The state x(t) of the
system determines the situation x(t + 1) at time t + 1. An example is is the Ulam map T (x) = 4x(1 − x) on
the interval [0, 1]. This is an example, where we have no idea what happens after a few hundred iterates even
if we would know the initial position with the accuracy of the Planck scale.
Dynamical system theory has applications all fields of mathematics. It can be used to find roots of equations
like for
T (x) = x − f (x)/f 0 (x) .
A system of number theoretical nature is the Collatz map
x
T (x) = (even x), 3x + 1 else .
2
A system of geometric nature is the Pedal map which assigns to a triangle the pedal triangle.
About 100 years ago, Henry Poincaré was able to deal with chaos of low dimensional systems. While
statistical mechanics had formalized the evolution of large systems with probabilistic methods already, the
new insight was that simple systems like a three body problem or a billiard map can produce very com-
plicated motion. It was Poincaré who saw that even for such low dimensional and completely deterministic
systems, random motion can emerge. While physisists have dealt with chaos earlier by assuming it or artifi-
cially feeding it into equations like the Boltzmann equation, the occurrence of stochastic motion in geodesic
flows or billiards or restricted three body problems was a surprise. These findings needed half a century to
sink in and only with the emergence of computers in the 1960ies, the awakening happened. Icons like Lorentz
helped to popularize the findings and we owe them the ”butterfly effect” picture: a wing of a butterfly can
produce a tornado in Texas in a few weeks. The reason for this statement is that the complicated equations
to simulate the weather reduce under extreme simplifications and truncations to a simple differential equation
ẋ = σ(y − x), ẏ = rx − y − xz, ż = xy − bz, the Lorenz system. For σ = 10, r = 28, b = 8/3, Ed Lorenz
discovered in 1963 an interesting long time behavior and an aperiodic ”attractor”. Ruelle-Takens called it a
123
strange attractor. It is a great moment in mathematics to realize that attractors of simple systems can
become fractals on which the motion is chaotic. It suggests that such behavior is abundant. What is chaos?
If a dynamical system shows sensitive dependence on initial conditions, we talk about chaos. We will
experiment with the two maps T (x) = 4x(1 − x) and S(x) = 4x − 4x2 which starting with the same initial
conditions will produce different outcomes after a couple of iterations.
The sensitive dependence on initial conditions is measured by how fast the derivative dT n of the n’th iterate
grows. The exponential growth rate γ is called the Lyapunov exponent. A small error of the size h will be
amplified to heγn after n iterates. In the case of the Logistic map with c = 4, the Lyapunov exponent is log(2)
and an error of 10−16 is amplified to 2n · 10−16 . For time n = 53 already the error is of the order 1. This
explains the above experiment with the different maps. The maps T (x) and S(x) round differently on the level
10−16 . After 53 iterations, these initial fluctuation errors have grown to a macroscopic size.
Here is a famous open problem which has resisted many attempts to solve it: Show that the map T (x, y) =
(c sin(2πx) + 2x − y, x) with T n (x, y) = (fn (x, y), gn (x, y)) has sensitive dependence on initial conditions on a
R1R1
set of positive area. More precisely, verify that for c > 2 and all n n1 0 0 log |∂x fn (x, y)| dxdy ≥ log( 2c ). The
left hand side converges to the average of the Lyapunov exponents which is in this case also the entropy of the
map. For some systems, one can compute the entropy. The logistic map with c = 4 for example, which is also
called the Ulam map, has entropy log(2). The cat map
T (x, y) = (2x + y, x + y) mod1
√
has positive entropy log |( 5 + 3)/2|. This is the logarithm of the larger eigenvalue of the matrix implementing
T.
While questions about simple maps look artificial at first, the mechanisms prevail in other systems: in astron-
omy, when studying planetary motion or electrons in the van Allen belt, in mechanics when studying coupled
pendulum or nonlinear oscillators, in fluid dynamics when studying vortex motion or turbulence, in geometry,
when studying the evolution of light on a surface, the change of weather or tsunamis in the ocean. Dynamical
systems theory started historically with the problem to understand the motion of planets. Newton realized
that this is governed by a differential equation, the n-body problem
n
X cij (xi − xj )
x00j (t) = ,
i=1
|xi − xj |3
where cij depends on the masses and the gravitational constant. If one body is the sun and no interaction of the
planets is assumed and using the common center of gravity as the origin, this reduces to the Kepler problem
x00 (t) = −Cx/|x|3 , where planets move on ellipses, the radius vector sweeps equal area in each time and the
period squared is proportional to the semi-major axes cubed. A great moment in astronomy was when Kepler
derived these laws empirically. An other great moment in mathematics is Newton’s theoretically derivation
from the differential equations.
124
OLIVER KNILL
Lecture 13: Computing

Computing deals with algorithms and the art of programming. While the subject intersects with computer sci-
ence, information technology, the theory is by nature very mathematical. But there are new aspects: computers
have opened the field of experimental mathematics and serve now as the laboratory for new mathematics.
Computers are not only able to simulate more and more of our physical world, they allow us to explore new
worlds.
A mathematician pioneering new grounds with computer experiments does similar work than an experimental
physicist. Computers have smeared the boundaries between physics and mathematics. According to Borwein
and Bailey, experimental mathematics consists of:
Gain insight and intuition. Explore possible new results
Find patterns and relations Suggest approaches for proofs
Display mathematical principles Automate lengthy hand derivations
Test and falsify conjectures Confirm already existing proofs
When using computers to prove things, reading and verifying the computer program is part of the proof. If
Goldbach’s conjecture would be known to be true for all n > 1018 , the conjecture should be accepted because
numerical verifications have been done until 2 · 1018 until today. The first famous theorem proven with the help
of a computer was the ”4 color theorem” in 1976. Here are some pointers in the history of computing:
2700BC Sumerian Abacus 1935 Zuse 1 programmable 1973 Windowed OS

200BC Chinese Abacus 1941 Zuse 3 1975 Altair 8800
150BC Astrolabe 1943 Harvard Mark I 1976 Cray I
125BC Antikythera 1944 Colossus 1977 Apple II
1300 Modern Abacus 1946 ENIAC 1981 Windows I
1400 Yupana 1947 Transistor 1983 IBM PC
1600 Slide rule 1948 Curta Gear Calculator 1984 Macintosh
1623 Schickard computer 1952 IBM 701 1985 Atari
1642 Pascal Calculator 1958 Integrated circuit 1988 Next
1672 Leibniz multiplier 1969 Arpanet 1989 HTTP
1801 Punch cards 1971 Microchip 1993 Web browser, PDA
1822 Difference Engine 1972 Email 1998 Google
1876 Mechanical integrator 1972 HP-35 calculator 2007 iPhone
We live in a time where technology explodes exponentially.Moore’s law from 1965 predicted that semiconductor
technology doubles in capacity and overall performance every 2 years. This has happened since. Futurologists
like Ray Kurzweil conclude from this technological singularity in which artificial intelligence might take over.
An important question is how to decide whether a computation is ”easy” or ”hard”. In 1937, Alan Turing
introduced the idea of a Turing machine, a theoretical model of a computer which allows to quantify com-
plexity. It has finitely many states S = {s1 , ..., sn , h } and works on an tape of 0 − 1 sequences. The state h is
the ”halt” state. If it is reached, the machine stops. The machine has rules which tells what it does if it is in
state s and reads a letter a. Depending on s and a, it writes 1 or 0 or moves the tape to the left or right and
moves into a new state. Turing showed that anything we know to compute today can be computed with Turing
machines. For any known machine, there is a polynomial p so that a computation done in k steps with that
computer can be done in p(k) steps on a Turing machine. What can actually be computed? Church’s thesis
of 1934 states that everything which can be computed can be computed with Turing machines. Similarly as in
mathematics itself, there are limitations of computing. Turing’s setup allowed him to enumerate all possible
Turing machine and use them as input of an other machine. Denote by T M the set of all pairs (T, x), where T
125
is a Turing machine and x is a finite input. Let H ⊂ T M denote the set of Turing machines (T, x) which halt
with the tape x as input. Turing looked at the decision problem: is there a machine which decides whether a
given machine (T, x) is in H or not. An ingenious Diagonal argument of Turing shows that the answer is ”no”.
[Proof: assume there is a machine HALT which returns from the input (T, x) the output HALT(T, x) = true,
if T halts with the input x and otherwise returns HALT(T, x) = false. Turing constructs a Turing machine
DIAGONAL, which does the following: 1) Read x. 2) Define Stop=HALT(x,x) 3) While Stop=True repeat
Stop:=True; 4) Stop.
Now, DIAGONAL is either in H or not. If DIAGONAL is in H, then the variable Stop is true which means
that the machine DIAGONAL runs for ever and DIAGONAL is not in H. But if DIAGONAL is not in H, then
the variable Stop is false which means that the loop 3) is never entered and the machine stops. The machine is
in H.]
Lets go back to the problem of distinguishing ”easy” and ”hard” problems: One calls P the class of decision
problems that are solvable in polynomial time and NP the class of decision problems which can efficiently be
tested if the solution is given. These categories do not depend on the computing model used. The question
”N=NP?” is the most important open problem in theoretical computer science. It is one of the seven mille-
nium problems and it is widely believed that P 6= N P . If a problem is such that every other NP problem
can be reduced to it, it is called NP-complete. Popular games like Minesweeper or Tetris are NP-complete. If
P 6= N P , then there is no efficient algorithm to beat the game. The intersection of NP-hard and NP is the class
of NP-complete problems. An example of an NP-complete problem is the balanced number partitioning
problem: given n positive integers, divide them into two subsets A, B, so that the sum in A and the sum in B
are as close as possible. A first shot: chose the largest remaining number and distribute it to alternatively to
the two sets.
We all feel that it is harder to find a solution to a problem rather than to verify a solution. If N 6= N P
there are one way functions, functions which are easy to compute but hard to verify. For some important prob-
lems, we do not even know whether they are in NP. Examples are the the integer factoring problem. An
efficient algorithm for the first one would have enormous consequences. Finally, lets look at some mathematical
problems in artificial intelligence AI:
problem solving playing games like chess, performing algorithms, solving puzzles
pattern matching speech, music, image, face, handwriting, plagiarism detection, spam
reconstruction tomography, city reconstruction, body scanning
research computer assisted proofs, discovering theorems, verifying proofs
data mining knowledge acquisition, knowledge organization, learning
translation language translation, porting applications to programming languages
creativity writing poems, jokes, novels, music pieces, painting, sculpture
simulation physics engines, evolution of bots, game development, aircraft design
inverse problems earth quake location, oil depository, tomography
prediction weather prediction, climate change, warming, epidemics, supplies
126
OLIVER KNILL
About this document

It should have become obvious that I’m reporting on many of these theorems as a tourist and
not as a local. In some few areas I could qualify as a tour guide but hardly as a local. The
references contain only parts consulted but it does not imply that I know all of that source. My
own background was in dynamical systems theory and mathematical physics. Both of these
subjects by nature have many connections with other branches of mathematics.
The motivation to try such a project came through teaching a course called Math E 320 at the
Harvard extension school. This math-multi-disciplinary course is part of the “math for teaching
program”, and tries to map out the major parts of mathematics and visit some selected placed
on 12 continents.
It is wonderful to visit other places and see connections. One can learn new things, relearn
old ones and marvel again about how large and diverse mathematics is but still to notice how
many similarities there are between seemingly remote areas. A goal of this project is also to
get back up to speed up to the level of a first year grad student (one forgets a lot of things over
the years) and maybe pass the quals (with some luck).
This summer 2018 project also illustrates the challenges when trying to tour the most important
mountain peaks in the mathematical landscape with limited time. Already the identification
of major peaks and attaching a “height” can be challenging. Which theorems are the most
important? Which are the most fundamental? Which theorems provide fertile seeds for new
theorems? I recently got asked by some students what I consider the most important theorem
in mathematics (my answer had been the “Atiyah-Singer theorem”).
Theorems are the entities which build up mathematics. Mathematical ideas show their merit
only through theorems. Theorems not only help to bring ideas to live, they in turn allow to
solve problems and justify the language or theory. But not only the results alone, also the
history and the connections with the mathematicians who created the results are fascinating.
The first version of this document got started in May 2018 and was posted in July 2018. Com-
ments, suggestions or corrections are welcome. I hope to be able to extend, update and clarify
it and explore also still neglected continents in the future if time permits.
It should be pretty obvious that one can hardly do justice to all mathematical fields and that
much more would be needed to cover the essentials. A more serious project would be to identify
a dozen theorems in each of the major MSC 2010 classification fields. This would roughly lead
to a “thousand and one theorem” list. In some sense, this exists already: on Wikipedia, there
are currently about 1000 theorems discussed. The one-document project getting closest to this
project is maybe the beautiful book [315].
127
173. Document history

The first draft was posted on July 22 [252]. On July 23, a short list of theorems was made
available on [253]. This document history section got started July 25-27, 2018.
• July 28: Entry 36 had been a repeated prime number theorem entry. Its alternative is
now the Fredholm alternative. Also added are the Sturm theorem and Smith normal
form.
• July 29: The two entries about Lidskii theorem and Radon transform are added.
• July 30: An entry about linear programming.
• July 31: An entry about random matrices.
• August 2: An entry about entropy of diffeomorphisms
• August 4: 104-108 entries: linearization, law of small numbers, Ramsey, Fractals and
Poincare duality.
• August 5: 109-111 entries: Rokhlin and Lax approximation, Sobolev embedding
• August 6: 112: Whitney embedding.
• August 8: 113-114: AI and Stokes entries
• August 12: 115 and 116: Moment entry and martingale theorem
• August 13: 117 and 118: theorema egregium and Shannon theorem
• August 14: 119 mountain pass
• August 15: 120, 121,122,123 exponential sums, sphere theorem, word problem and finite
simple groups
• August 16: 124, 125, 126, Rubik, Sard and Elliptic curves,
• August 17: 127, 128, 129 billiards, uniformization, Kalman filter
• August 18: 130,131 Zarisky and Poincare’s last theorem
• August 19: 132, 133 Geometrization, Steinitz
• August 21: 134, 135 Hilbert-Einstein, Hall marriage
• August 22: 136-130
• August 24: 141-142
• August 25: 143-144
• August 27: 145-149
• August 28: 150-151
• August 31: 152
• September 1: 153-155
• September 2: 156
• September 8: 157,158
• September 14: 159-161
• September 25 2018: 162-164
• March 17 2019: 165-169
• March 20, 2019: section on paradigms
• March 21, 2019: 170
• March 27, 2019, 171
• June 20, 2019, 172
128
OLIVER KNILL
174. Top choice

The short list of 10 theorems mentioned in the youtube clip were:
• Fundamental theorem of arithmetic (prime factorization)
• Fundamental theorem of geometry (Pythagoras theorem)
• Fundamental theorem of logic (incompleteness theorem)
• Fundamental theorem of topology (rule of product)
• Fundamental theorem of computability (Turing computability)
• Fundamental theorem of calculus (Stokes theorem)
• Fundamental theorem of combinatorics, (pigeon hole principle)
• Fundamental theorem of analysis (spectral theorem)
• Fundamental theorem of algebra (polynomial factorization)
• Fundamental theorem of probability (central limit theorem)
Let us try to justify this shortlist. It should go without saying that similar arguments could be
found for any other choice except maybe for the five classical fundamental theorems: Arithmetic,
Geometry (which is undisputed Pythagoras), Calculus and Algebra, where one can hardly must
argue much: except for Pythagoras, their given name already suggests that they are considered
fundamental.
• Analysis. Why chose the spectral theorem and not say the more general Jordan nor-
mal form theorem? This is not an easy call but the Jordan normal form theorem
is less simple to state and furthermore, that it does not stress the importance of nor-
mality giving the possibility for a functional calculus. Also, the spectral theorem
holds in infinite dimensions for operators on Hilbert spaces. If one looks at mathemat-
ical physics for example, then it is the functional calculus of operators which is
really made use of; the Jordan normal form theorem appears rarely in comparison. In
infinite dimensions, a Jordan normal form theorem would be much more difficult as the
operator Au(n) = u(n + 1) on l2 (Z) is both unitary as well as a “Jordan form matrix”.
The spectral theorem however sails through smoothly to infinite dimensions and even
applies with adaptations to unbounded self-adjoint operators which are important
in physics. And as it is a core part of analysis, it is also fine to see the theorem as part
of analysis. The main reason of course is that the fundamental theorem of algebra is
already occupied by a theorem. One could object that “analysis” is already represented
by the fundamental theorem of calculus but calculus is so important that it can represent
its own field. The idea of the fundamental theorem of calculus goes beyond calculus. It
is essentially a cancellation property, a telescopic sum or Pauli principle (d2 = 0
for exterior derivatives) which makes the principle work. Calculus is the idea of an
exterior derivative, the idea of cohomology, a link between algebra and geometry. One
can see calculus also as a theory of “time”. In some sense, the fundamental theorem of
calculus also represents the field of differential equations and this is what “time is
all about”.
• Probability. One can ask also why to pick the central limit theorem and not say
the Bayes formula or then the deeper law of iterated logarithm. One objection
against the Bayes formula is that it is essentially a definition, like the basic arithmetic
properties “commutativity, distributivity or associativity” in an algebraic structure like
129
a ring. One does not present the identity a + b = b + a for example as a fundamental
theorem. Yes, the Bayes theorem has an unusual high appeal to scientists as it appears
like a magic bullet, but for a mathematician, the statement just does not have enough
beef: it is a definition, not a theorem. Not to belittle the Bayes theorem, like the notion
of entropy or the notion of logarithm, it is a genius concept. But it is not an
actual theorem, as the cleverness of the statement of Bayes lies in the definition and
so the clarification of conditional probability theory. For the central limit theorem, it is
pretty clear that it should be high up on any list of theorems, as the name suggests: it is
central. But also, it actually is stronger than some versions of the law of large numbers.
The strong law is also super seeded by Birkhoff’s ergodic theorem which is much more
general. One could argue to pick the law of iterated logarithm or some Martingale
theorem instead but there is something appealing in the central limit theorem which
goes over to other set-ups. One can formulate the central limit theorem also for random
variables taking values in a compact topological group like when doing statistics with
spherical data [316]. An other pitch for the central limit theorem is that it is a fixed
point of a renormalization map X → X + X (where the right hand side is the sum
of two independent copies of X) in the space of random variables. This map increases
entropy and the fixed Rpoint is is a random variable whose distribution function f has the
maximal entropy − R f (x) log(f (x)) dx among all probability density functions. The
entropy principle justifies essentially all known probability density functions. Nature
just likes to maximize entropy and minimize energy or more generally - in the presence
of energy - to minimize the free energy.
• Topology. Topology is about geometric properties which do not change under contin-
uous deformation or more generally under homotopies. Quantities which are invariant
under homeomorphisms are interesting. Such quantities should add up under disjoint
unions of geometries and multiply under products. The Euler characteristic is the proto-
type. Taking products is fundamental for building up Euclidean spaces (also over other
fields, not only the real numbers) which locally patch up more complicated spaces. It is
the essence of vector spaces that after building a basis, one has a product of Euclidean
spaces. Field extensions can be seen therefore as product spaces. How does the counting
principle come in? As stated, it actually is quite strong and calling it a “fundamental
principle of topology” can be justified if the product of topological spaces is defined
properly: if 1 is the one-point space, one can see the statement G × 1 = G1 as the
Barycentric refinement of G, implying that the Euler characteristic is a Barycentric
invariant and so that it is a “counting tool” which can be pushed to the continuum, to
manifolds or varieties. And the compatibility with the product is the key to make it
work. Counting in the form of Euler characteristic goes throughout mathematics, com-
binatorics, differential geometry or algebraic geometry. Riemann-Roch or Atiyah-Singer
and even dynamical versions like the Lefschetz fixed point theorem (which generalizes
the Brouwer fixed point theorem) or the even more general Atiyah-Bott theorem can be
seen as extending the basic counting principle: the Lefschetz number χ(X, T )
is a dynamical Euler characteristic which in the static case T = Id reduces to the Euler
characteristic χ(X). In “school mathematics”, one calls the principle the “fundamental
130
OLIVER KNILL
principle of counting” or “rule of product”. It is put in the following way: “If we have
k ways to do one thing and m ways to do an other thing, then we have k ∗ m ways to
do both”. It is so simple that one can argue that it is over represented in teaching but
it is indeed important. [44] makes the point that it should be considered a founding
stone of combinatorics.
Why is the multiplicative property more fundamental than the additive counting
principle. It is again that the additive property is essentially placed in as a definition of
what a valuation is. It is in the in-out-formula χ(A ∪ B) + χ(A ∩ B) = χ(A) + χ(B).
Now, this inclusion-exclusion formula is also important in combinatorics but it is already
in the definition of what we call counting or “adding things up”. The multiplicative
property on the other hand is not a definition; it actually is quite non-trivial. It charac-
terizes classical mathematics as quantum mechanics or non-commutative flavors
of mathematics have shown that one can extend things. So, if the “rule of product”
(which is taught in elementary school) is beefed up to be more geometric and interpreted
to Euler characteristic, it becomes fundamental.
• Combinatorics. The pigeon hole principle stresses the importance of order struc-
ture, partially ordered sets (posets) and cardinality or comparisons of cardinality. The
point for posets is made in [336] who writes The biggest lesson I learned from Richard
Stanley’s work is, combinatorial objects want to be partially ordered! The use of injec-
tive functions to express cardinality is a key part of Cantor. Like some of the ideas
of Grothendieck it is of “infantile simplicity” (quote Grothendieck about schemes) but
powerful. It allowed for the stunning result that there are different infinities. One of
the reason for the success of Cantor’s set theory is the immediate applicability. For any
new theory, one has to ask: “does it tell me something I did not know?” In “set theory”
the larger cardinality of the reals (uncountable) than the cardinality of the algebraic
numbers (countable) gave immediately the existence of transcendental numbers.
This is very elegant. The pigeon hole principle similarly gives combinatorial results
which are non trivial and elegant. Currently, searching for “the fundamental theorem
of combinatorics” gives the “rule of product”. As explained above, we gave it a geo-
metric spin and placed it into topology. Now, combinatorics and topology have always
been very hard to distinguish. Euler, who somehow booted up topology by reducing
the Königsberg problem to a problem in graph theory did that already. Combi-
natorial topology is essentially part of topology. Today, some very geometric topics
like algebraic geometry have been placed within pure commutative algebra (this is
how I myself was exposed to algebraic geometry) On the other hand, some very hard
core combinatorial problems like the upper bound conjecture have been proven with
algebro-geometric methods like toric varieties which are geometric. In any case, order
structures are important everywhere and the pigeon principle justifies the importance
of order structures.
• Computation. There is no official “fundamental theorem of computer science” but
the Turing completeness theorem comes up as a top candidate when searching on
engines. Turing formalized using Turing machines in a precise way, what computing
is, and even what a proof is. It nails down mathematical activity of running an
131
algorithm or argument in a mathematical way. It is also pure as it is not hardware

dependent. One can also only appreciate Turing’s definition if one sees how different
programming languages can look like and also in logic, what type of different frame
works have been invented. Turing breaks all this complexity with a machine which can
be itself part of mathematics leading to the Halte problem illustrating the basic limi-
tations of computation. Quantum computing would add a hardware component and
might break through the Turing-Church thesis that everything we can compute can
be computed with Turing machines in the same complexity class. Goedel and Turing
are related and the Turing incompleteness theorem has a similar flavor than the Goedel
incompleteness theorems. There is an other angle to it and that is the question of com-
plexity. I would predict that most mathematicians would currently favor the Platonic
view of the Church thesis and predict that also new paradigms like quantum comput-
ing will never go beyond Turing computability or even not break through complexity
barriers like P-NP thresholds. It is just that the Turing completeness theorem is too
beautiful to be spoiled by a different type of complexity tied to a physical world. The
point of view is that anything we see in the physical world can in principle be computed
with a machine without changing the complexity class. But that picture could
be as naive as Hilbert’s dream one hundred years ago. Still, whatever happens in the
future, the Turing completeness theorem remains a theorem. Theorems stay true.
• Logic. One can certainly argue whether it would be justified to have Goedel’s theorem
replaced by a theorem in category theory like the Yoneda lemma. The Yoneda result is
not easy to state and it does not produce yet an “Aha moment” like Goedel’s theorem
does (the liars paradox explains the core of Goedel’s theorem). Maybe it will so in
the future, when all mathematics has been naturally and pedagogically well expressed
in categorical language. I’m personally not sure whether this will ever happen: not
everything which is nice also had been penetrating large parts of mathematics: an ex-
ample is given by non-standard analysis, which makes calculus orders of magnitudes
easier and which is related also to surreal numbers, which are the most “natural”
numbers. Both concepts have not entered calculus or algebra textbooks and there are
reasons: the subjects need mathematical maturity and one can easily make mistakes. (I
myself use non-standard analysis on an intuitive level as presented by Nelson [314, 345]
and think of a compact set as a finite set for example which for example, where basic
theorems almost require no proof like the Bolzano theorem telling that a continuous
function on a compact set takes a maximum). But using non-standard analysis would
be a “no-no” both in teaching as well when formulating mathematical thoughts for oth-
ers who are not familar with the three additional axioms IST within ZFC of Nelson. It
is non-standard and true to its name. Also, much of category theory still feels just like
a huge conglomerate of definitions. There is lots of dough in the form of definitions and
little raisins in the form of theorems. Historically also the language of set theory have
been overkill especially in education, where it has lead to “new math” controversies in
the 1960ies. The work of Russel and Whitehead demonstrates, how clumsy things can
become if boiled down to the small pieces. We humans like to think and programming
in higher order structures, rather than doing assembly coding, we like to work in object
132
OLIVER KNILL
oriented languages which give more insight. But we like and make use of that higher
order codes can be boiled down to assembly closer to what the basic instructions are.
This is similar in mathematics and also in future, a topologist working in 4 manifold
theory will hardly think about all the definitions in terms of sets for similar reasons than
a modern computer algebra system does not break down all the objects into lists and
lists of lists (even so, that’s what it often is). Category theory has a chance to change
the landscape because it is close to computer science and to natural data structures. It
is more pictorial and flexible than set theory alone. It definitely has been very successful
to find new structures and see connections within different fields like computer science
[331]. It also has lead to more flexible axiom systems
133
Index
E8 lattice, 41 associativity, 7, 21
σ-algebra, 4 Atiyah Bott, 87
15 theorem, 41 Atiyah Singer, 87
17-gon, 38 Atiyah-Singer theorem, 127
290 theorem, 41 attractor of iterated function system, 45
3-connected, 62 Aubry-Mather theory, 30, 36
3-sphere, 61 Axiom of choice, 72
3D scanning, 95 axiom of choice, 5
4 color theorem, 65 axiom system, 8, 9
4-color theorem, 87
Bézier curves, 27
ABC conjecture, 88 Bézout’s bound, 20
abscissa of absolute convergence, 12 Baire category theorem, 18
abscissa of convergence, 12 Baire space, 75
absolute Galois group, 23 Bakshali manuscript, 104
absolutely continuous measure, 14 Banach algebra, 10
adjoint, 6 Banach fixed point theorem, 12
affine camera, 31 Banach space, 8, 38, 64
Alaoglu theorem, 30 Banach-Tarski construction, 70
Alexander polynomial, 32 Banach-Tarski paradox, 92
Alexander sphere, 92 Barycenter, 38
algebraic closure, 23 Barycentric subdivision, 11
algebraic extension, 24 Bayes theorem, 4
Algebraic number, 27 Beals conjecture, 108
algebraic number field, 23 Bernoulli shift, 22
algebraic numbers, 23 Bernstein polynomials, 27
Algebraic set, 5 Bertrand postulate, 68
algorithm, 95 Bertrand’s theorem, 74
Almost complex structure, 18 Betti number, 17
almost everywhere convergence, 70 bifurcation, 11, 75
Almost Mathieu operator, 92 bijective, 3
alphabet, 22 billiards, 58
Alternating sign conjecture, 79 Binomial coefficients, 27
Alternating sign matrix, 79 Bipartite graph, 81
AMS classification, 92, 102 bipartite graph, 63
analytic function, 7 Birkhoff, 3
Analytical index, 37 Birkhoff theorem, 3
ancient roots, 101 Boltzmann constant, 53
Angle trisector, 77 Bolzano-Weierstrass theorem, 78
Angular momentum, 74 Bolzmann equation, 30
Anosov, 28 Borel measure, 9, 14, 51
antipodal point, 38 Borsuk-Ulam theorem, 38
aperiodic, 47 boundary, 49
area element, 35 bounded linear operator, 6
area of polygon, 29 bounded martingale, 52
area triangle, 35 bounded stochastic process, 52
argument, 13 Brauer group, 33
arithmetic mean, 19 Brioschi formula, 53
arithmetic progression, 15 Brjuno set, 27
artificial intelligence, 95 Brouwer degree, 25
Arzela-Ascoli, 39 Brouwer fixed point theorem, 14, 25
asset pricing, 52 Burnside lemma, 28
134
OLIVER KNILL
Butterfly theorem, 105 cohomology, 25

coloring, 46, 65
C star algebra, 17 Combinatorial convexity, 80
Césaro convergence, 19 combinatorics, 115
Cahen formula, 12 commutative, 7
Calculus of variations, 79 commutative ring, 5, 49
calculus of variations, 93 compact group, 9
canonical divisor, 30 Compact operator, 15
canons of rhetorik, 101 compact-open topology, 40
Cantor, 3
compactly supported, 18
Cantor set, 45, 92
compactness condition, 54
Caratheodory, 3
complementary angles, 35
cardinality, 3, 7, 11, 72
complete, 8, 40
Carleson theorem, 71
Complete metric space, 18
Cartesian closed, 22
complete metric space, 11
Cartesian product, 11
completeness, 12
catastrophe, 75
complex conjugate, 6
catastrophe theory, 81
complex logarithm, 13
category, 10, 22
complex manifold, 59
Cauchy, 14
complex multiplication, 38
Cauchy integral formula, 8
complex plane, 7
Cauchy integral theorem, 8
complexity theory, 49
Cauchy sequence, 11
Cauchy-Kovalevskaya theorem, 14 composition of functions, 34
Cauchy-Riemann, 7 computer algebra, 20
Cavalieri principle, 29 computer assisted proof, 39
Cayley theorem, 10 computer vision, 49
cellular automaton, 22 conditional expectation, 52
center of mass, 18 conditional probability, 4
central force, 74 conductor, 38
central limit theorem, 2, 116 conformally equivalent, 59
centrally symmetric, 15 conic, 69
Ceva theorem, 76 conic section, 20, 69
chamber, 118 connection, 16
character, 9, 10 Connes formula, 17
Chebyshev inequality, 89 conservation law, 93
Chern character, 37 consistent, 8
Chern class, 68 context, 95
Chern classes, 37 continued faction expansion, 36
Chinese Remainder Theorem, 7 continuous function, 26
chiral copies, 40 continuous functions, 34
chord tangent construction, 58 continuous map, 25
Church, 10 continuously differentiable, 2
Church thesis, 10 Continuum hypothesis, 72
Church-Turing thesis, 10 contractible, 5
circle, 7 contraction, 11
Circular law, 44 convergence, 12
circumference of circle, 13 convergence in distribution, 2
class fields, 38 convex, 15, 19, 38
Classification of finite simple groups, 88 convex function, 19
classification of finite simple groups, 56 convex polytop, 43, 66
close orbits, 28 convex set, 19
closed manifold, 61 convolution, 10
cognitive science, 49 Conway-Schneeberger fifteen theorem, 41
Cohomology, 47 coprime, 6, 20
135
cost function, 30 Dirac delta function, 27

countable, 3 Dirac operator, 17, 47
coupling transformation, 30 direct sum of representations, 9
cover, 49 Dirichlet, 7
critical point, 17, 54, 79 Dirichlet Eta function, 12
critical points, 57 Dirichlet prime number theorem, 15
critical set, 57 Dirichlet problem, 73
CRT system, 20 Dirichlet series, 12
crypto system, 6 Dirichlet unit problem, 36
crystallogrphy, 40 discrete σ algebra, 70
cube, 17 discrete dynamical system, 60
cube exchange transformation, 48 discrete Ito integral, 52
cubic equation, 24 discrete log, 13
cumulative distribution function, 2 discrete log problem, 6
Curtis-Hedlund-Lyndon, 22 discrete time stochastic process, 52
curvature, 36 discriminant, 53
cyclic cube exchange transformations, 48 Disordered system, 76
cyclic polytop, 66 divergence, 21
cyclic subgroup, 21 division algebra, 33, 88
Cyclotomic field, 38 divisor, 30
cyclotomic field, 37 dodecahedron, 17
doformation idea, 90
Darboux, 16 Dolbeault operators, 69
data fitting, 8, 31 Doob martingale convergence, 52
data mining, 49 dual linear programming problem, 43
data science, 49 Dual numbers, 33
de la Vallee-Poussin, 21 dual numbers, 33
De Moivre, 3 duality, 90
Deformation idea, 90 dynamical system, 75
degree, 2
degree of a divisor, 30 edge labeling, 46
Dehn-Sommerville conditions, 67 edges, 4, 46
Demokrit, 93 Edwards-Anderson model, 76
density of sphere packing, 41 effective divisor, 30
Depressed cubic, 24 eigenbasis, 6
derivative, 2 Eigenvalue, 31
determinant, 16 Eigenvalues, 73
Diagonalizable, 6 eigenvalues, 6, 44
diagonalization, 13 Einstein constant, 62
diffeomorphic, 49 Einstein formula, 86
Diffeomorphism, 28 elementary catastrophe, 75
diffeomorphism, 68 elementary theory of the category of sets, 9
differentiable, 2 elliptic curve, 20, 57
differentiable manifold, 17 Elliptic curve cryptography, 58
differentiable sphere theorem, 55 elliptic regularity, 31, 37
differential equation, 8 embedding, 19, 49
differential form, 50 Embedding theorems, 89
Differential operator, 29, 31 empty set, 3
differential operator, 37, 48 energy momentum tensor, 62
Diffie-Hellman system, 6 Entropy, 44
dimension, 11 entropy, 53, 58, 116
Diophantie approximation, 92 equalizer, 22
Diophantine, 27 equicontinuous, 39
Diophantine analysis, 15 equilibrium, 14
Diophantine number, 27, 35 equilibrium point, 11
136
OLIVER KNILL
equivalence class, 21 Fourier basis, 19

Erathostenes, 93 Fourier coefficients, 19, 27
Erdös conjecture on arithmetic progressions, 16 Fourier series, 19, 43, 70
Ergodic, 3 Fourier theory, 2
ergodic, 39 Fourier transform, 10
Erlangen program, 105 Fréchet derivative, 54
error correction, 122 Fréchet space, 64
Euclidean algorithm, 20 fractal, 45
Euler characteristic, 4, 5, 11, 88 Fredholm alternative, 15
Euler gem formula, 5 Fredholm theory, 15, 31
Euler golden key, 12, 21, 86 Free action, 47
Euler handshake, 4 Friendship problem, 46
Euler polyhedron formula, 5 Frobenius determinant, 80
Euler polynomial, 11 Frobenius determinant theorem, 80
Euler product, 21 functional calculus, 6
Euler totient function, 6 functor, 10
Euler-Lagrange equations, 18, 79 fundamental class, 37
exhaustion, 13 fundamental counting principle, 11
Exotic sphere, 92 fundamental group, 72
exotic sphere, 55 fundamental lemma of calculus of variations, 18
expectation, 116 fundamental region, 15
exponential map, 13 fundamental theorem of algebra, 2
exponential of categories, 22 fundamental theorem of Riemannian geometry, 16
extending mathematical operations, 103 fundamental theorem of trigonometry, 13
exterior algebra, 33 fundamental theorem of Vlasov dynamics, 33
exterior derivative, 17
extreme point, 38, 66 g-conjecture, 67
Galois extension, 24
f-vector, 66 Galois theory, 90
Fürstenberg theorem, 23 Game of life, 22
Féjer kernel, 19 Gauss, 1, 21
faces, 62 Gauss Bonnet, 80
Factorial, 70 Gauss identity, 86
factorization domain, 1 Gauss sums, 54
Feigenbaum conjectures, 87 Gauss-Bonnet, 4, 87
Feigenbaum universality, 39 Gauss-Bonnet-Chern, 36
Feigenbaum-Civtanovic functional equation, 39 Gaussian curvature, 36
Fermat law, 79 Gaussian random variables, 2
Fermat prime, 38 general position, 20
Fermat’s last theorem, 38 general recursive function, 10
Fermat’s little theorem, 6 General theory of relativity, 63
Fermat’s principle, 79 generalized functions, 48
Feuerbach theorem, 105 Generalized handshake, 4
field extension, 24 Generalized Poincare conjecture, 61
fifteen theorem, 41 Generating function, 28
finite group, 9, 21 Generic, 18
finite simple graph, 4 genus, 30
finitely presented group, 55 geodesic complete, 40
first Baire category, 18 geodesic distance, 17, 40
First fundamental form, 52 geometric mean, 19
first integral, 18 geometric realization, 62
fixed point set, 28 Gibbs formula, 86
fixed points, 25 Girko law, 44
formal rules, 8 global maximum, 79
four square theorem, 26, 32 God number, 110
137
god number, 57 Holder continuity, 48

Goedel, 9 Hollywood, 88
Goldbach conjecture, 68, 95 holomorphic, 40
Golden ratio, 92 holomorphic function, 7
Grönwall inequality, 44 homeomorphism, 23, 25
graph, 4 HOMFLYPT, 32
graph coloring, 65 homogeneous polynomial, 20
graph database, 95 homology group, 72
Green Tao Theorem, 15 homotopica, 61
Green’s formula, 73 homotopy group, 72
Green’s function, 31 homotopy idea, 90
Gregory, 2 Hook’s law, 74
Grobman-Hartman linearization, 44 Hopf method, 59
Grothendick group completion, 7 Hopf Umlaufsatz, 80
Grothendieck, 7 Hopf-Rinov theorem, 40
Grothendieck program, 90 horse shoe, 28
Group, 28 Huntchingson operator, 45
group, 7, 21 Hurewicz homomorphism, 73
group completion, 7 Hurewicz theorem, 72
Guss-Bonnet formula, 35 Hurwitz estimate, 89
hyperbola, 69
h-vector, 66 hyperbolic geometry, 35
Haar measure, 9, 10, 70 hyperbolic set, 28
Hadamard, 21 hypercomplex algebra, 33
Hadwiger, 13 hypercomplex numbers, 33
Hahn-Banach, 5 hypersurface, 20
Hall Mariage, 48
Hall marriage problem, 63 icosahedron, 17
Hamilton principle, 79 ideal, 5
happy groups, 56 identity element, 7
Harmitian, 42 ill posed problem, 43
harmonic forms, 69 image, 8
harmonic series, 21 imaginary quadratic field, 38
Hausdorff dimension, 27, 45 implicit function theorem, 11, 35
Hausdorff dimension of measure, 44 inconsistent, 8
Hausdorff metric, 13 independence, 53
Hausdorff space, 40 independent random variable, 2
Hedlund-Curtis-Lyndon, 22 index, 8, 21
Heine-Cantor theorem, 78 Inequalities, 89
Heisenberg model, 76 initial value theorem, 8
Hilbert, 6 injective, 3
Hilbert action, 63 inner product, 2, 16, 19
Hilbert cube, 88 Inner Universal Teichmuller Theory, 88
Hilbert Schmidt kernel, 15 Inscribed angle theorem, 76
Hilbert space, 6, 17 integer polynomial, 23
Hilbert’s 12th problem, 38 integer quadratic form, 41
Hilbert’s problems, 72 integrable, 39
Hilbert’s program, 9 integrable function, 19
Hilbert-Einstein equations, 63 integral, 2
Hilbert-Waring theorem, 31 integral operator, 15
Hippocrates Theorem, 78 integral quadratic form, 41
Hippocrates theorem, 78 invariant tori, 39
Hodge dual, 47 invariant valuation, 13
Hodge operator, 17 inverse, 7
Hofstadter butterly, 120 inverse problem, 43, 95
138
OLIVER KNILL
inverse problems, 49 lattice, 15

irreducible representation, 9 lattice gas model, 75
Ishango bone, 103 law of iterated logarithm, 116
Ising model, 75 law of large numbers, 3, 116
isoperimetric inequality, 29 law of quadratic reciprocity, 25
isospectral deformation, 18 Lax system, 18
Isospectral drum, 73 least square solution, 8
iterated function system, 45 Lebesgue measure, 15, 70
Lebowski theorem, 50
Jacobi matrix, 36 Ledrappier-Young formula, 44
Jacobi triple product, 28 Leech lattice, 41
Jensen inequality, 19 Lefschetz fixed point theorem, 14
Jones polynonial, 32 Lefschetz-Hopf fixed point theorem, 25
Jordan block, 78 left invariant, 9
Jordan curve, 19 Legendre conjecture, 68
Jordan normal form, 78 Legendre symbol, 25
Jordan-Brouwer separation theorem, 20 Leibniz, 2
Jordan-Chevalley decomposition, 78 Leibniz rule, 16
Julia set, 26 length of polygon, 29
Lerch transcendent, 12
K-Theory, 7 Levi Civita, 16
Kähler class, 68 Levi-Civita connection, 16
Kähler manifold, 68 Levy, 3
Kakutani fixed point theorem, 14 liberal arts and sciences, 101
Kakutani skyscraper, 47 Lidelof, 8
Kalman filter, 60 Lidskii-Last, 42
KAM, 36, 87 Lie group, 13
KAM theory, 30, 35 limit, 11, 12
KAM tori, 92 limit cycle, 26
Kepler problem, 87 line bundle, 37
kernel, 8 line segment, 19
Klein Erlangen program, 90 linear, 2
Klein model, 35 linear program, 43
knot, 32, 76 link, 32
knot sum, 32 Liouville integrable, 18
Koch curve, 45 Lipschitz, 8
Koenigsberg bridge problem, 117 local automaton, 22
Koopman theory, 90 local continuation, 10
Kowalevskaya, 14 local field, 38
Kronecker pairing, 37 local maximum, 79
Kronecker-Weber theorem, 37 locally compact, 9
Kuratowski, 5 locally compact topological space, 38
Locally convex, 64
L’Hopital rule, 13 logarithm, 13, 21
Lagrange equations, 29, 79 Logistic map, 92
Lagrange four square theorem, 41 Lorenz attractor, 26
Lagrange rest term, 29 lunes of Hippocrates, 78
Lagrange theorem, 21 Lyapunov exponent, 44, 58
Landau conjecture, 68
Landau problem, 16, 88, 95 Magnetic resonance imaging, 42
Landsberg-Schaar relation, 54 Mandelbrot set, 26, 63, 92
Langlands program, 88 Mandelbulb set, 63, 92
Laplace equation, 31 manifold, 17, 49
Laplacian, 73 Margulis-Ruelle inequality, 44
last theorem of Fermat, 87 Markov matrix, 71
139
marriage condition, 63 multi-complex, 67

martingale, 52 Multibrot set, 26
martingale transform, 52 multiple connections, 4
Matching, 81 multiple recurrent, 23
Math in movies, 88 multivariate moments, 51
mathematical roots, 101
Mather set, 92 n-body problem, 33
Mathieu operator, 120 n-connected space, 73
matrix, 8 n-linearization, 45
maximal equilibrium, 14 Nash equilibrim, 14
maximum principle, 43, 89 Nash-Kuiper, 92
Maxwell equations, 86 natural logarithm, 13, 21
Mazur torsion theorem, 65 natural numbers, 1
meager set, 18 natural transformation, 10
meagre set, 18 new foundations, 9
mean, 2 Newton, 2
Mean value theorem, 29 Newton method, 11
measure preserving, 3 Newton potential, 74
memory, 95 Nine chapters on the Mathematical art, 104
Noether theorem, 93
Menger carpet, 45
non-commutative geometry, 17, 93
Menger sponge, 92
non-degenerate, 10
Mergelyan theorem, 27
non-degenerate 2-form, 16
meromorphic function, 30
non-degenerate critical point, 17
Mersenne primes, 107
non-Euclidean geometry, 35
method of characteristics, 33
non-negative matrix, 71
metric space, 11
non-singular curve, 57
metrizable, 39
normal extension, 24
Millenium problems, 88
normal family, 40
mimiocretin, 50
normal group, 56
Minimal Matching, 81
normal operator, 6
Minkowski, 15
normal subgroup, 55
mixing, 39
notation, 95
moduli, 20
nuclear magnetic resonance, 42
moduli space, 105
Nullstellensatz, 5
moment generating function, 51
Number theory, 15
Moment methods, 51 Numerical analysis, 27
Monge-Kantorovich, 30 numerical methods, 93
monique polynomial, 24
monochromatic, 46 OCR, 49
monoid, 7 octahedron, 17
monomial, 20 octic conjecture, 34
monster, 56 octonions, 33
Montel theorem, 40 open domain, 7
Mordell-Weil theorem, 65 Open problems, 88
Morera’s theorem, 8 open set, 25
Morley theorem, 105 optimal sphere packing, 41
Morley triangle, 77 optimal stopping time, 52
morphism, 10 optimal transport, 30
Morse function, 17 Optimal transport problem, 30
Morse index, 17 optimization problem, 49
Morse inequality, 17 orbit, 28
Moser trick, 16 orthogonal complement, 8
mountain pass, 54 orthogonal group, 9
MRI, 42 Orthogonal projection, 31
140
OLIVER KNILL
orthonormal eigenbasis, 6 Poisson equation, 31

Ostrowski, 2 Poisson summation, 27
Ostrowski theorem, 88 Polish, 30
Polish space, 18, 88
p-adic field, 38 polychora, 118
p-adic integers, 9 polydisk, 29
p-adic valuation, 88 polyhedral formula, 62
Page rank, 81 polylogarithm, 12
page rank, 71, 81 polynomial, 2, 26
Palais-Smale condition, 54 polytopes, 118
Pappus theorem, 105 Pontryagin duality, 10
Paradigm, 91 positive definite tensor, 16
parametrization, 49 positive definition, 41
pariah, 56 positive matrix, 71
Parseval’s identity, 70 Potts model, 76
partial derivatives, 48 power set, 3
partial differences, 51 pre-sheave, 10
partial differential equation, 14 primality test, 70
partition, 21 prime, 6, 21
partition function, 86 prime counting function, 21, 68
Pascal configurations, 69 prime factorization, 1
Pascal theorem, 69 prime factors, 1
Patterns, 93 prime manifolds, 61
payoff function, 14 Prime number theorem, 21
Peano, 8 Prime twin conjecture, 95
Peano axioms, 8 primes, 1
pedagogy, 95 primitive root of unity, 37
Pedal triangle map, 123 principal curvature, 36
Pell’s equation, 36 principal divisor, 30
Pentagonal number theorem, 28 principal ideal domain, 42
Perelman theorem, 87 principal logarithm, 13
perfect field, 78 prizes, 88
periodic points, 47 probability measure, 4
perpendicular, 2, 8 probability space, 3, 4, 19
Perron-Frobenius theorem, 71 product topology, 5
Perseval equality, 2 profinite group, 9
Pesin formula, 44 projection function, 5
Pfaffian, 36 projective completion, 20
Phase transition, 89 projective geometry, 69
phase transition, 75 projective space, 20
Picard, 8 provably consistent, 8
Pick theorem, 73 pseudo orbit, 28
pigeon hole principle, 7 Puiseux formula, 53
planar, 62 pure point measure, 14
planar graph, 65 Pythagoras theorem, 2
Platonic polytop, 17
player, 14 quadratic form, 41
Poincaré Bendixon, 26 quadratic non-residue, 25
Poincaré conjecture, 90 quadratic reciprocity, 86
Poincaré disk, 35 quadratic residue, 25, 26
Poincaré recurrence, 23 quadrisecant, 76
Poincaré-Sigel theorem, 35 quadrivium, 101
Poincare conjecture, 61 quality, 88
Poincare duality, 47 quantisation, 93
Poisson commute, 18 quarter pinched, 55
141
quartic equation, 24 RSA crypto system, 6

quasi-linear Cauchy problem, 14 Rubik cube, 57, 110
Quasiperiodic, 39 Rubik cuboid, 57
quotient group, 21 Rudin sphere, 92
rule of product, 11
radical, 5, 88 Runge Kutta, 93
Radon theorem, 38, 80 Runge theorem, 27
Radon transform, 42, 43
Radon-Nikodym theorem, 15 saddle point, 54
Ramanujan primes, 68 Sard theorem, 57
Ramsey theory, 46 Schauder fixed point theorem, 25
random matrix, 43, 44 scheme, 49
random variable, 3 Schoenflies theorem, 20
rank-nullity, 8 Schröder equation, 35
real analytic, 28 Schroeder equation, 36
recursive function, 10 Schroeder-Bernstein, 3
reflection, 40 Schroedinger equation, 31
reflection ambiguity, 31 Schroedinger operator, 31, 36
regular Hausdorff, 25 Schwarz lemma, 89
regular n-gon, 38 Sclar curvature, 62
regular polygon, 17 Second Baire category, 18
relative density, 16 second countable, 25
reminder, 6 second fundamental form, 52
representation, 9 sectional curvature, 55
residual, 75 sedenions, 33
residual set, 18 Self adjoint, 42
residue, 8 self-adjoint operator, 6
Residue calculus, 8 self-loops, 4
resonance condition, 45 semi simple representation, 9
Revolution, 91 set category, 10
Ricci curvature, 68 sextic conjecture, 34
Ricci flow, 55 shadowing, 87
Ricci tensor, 62 shadowing property, 28
Riemann curvature tensor, 36 Shannon entropy, 53, 88
Riemann hypothesis, 88 Sherrington-Kirpatrick, 76
Riemann Lebesgue theorem, 27 Siegel linearization theorem, 35
Riemann mapping theorem, 59 Siegel theorem, 27
Riemann surface, 30, 59 sign changes, 42
Riemann zeta function, 21 similarity dimension, 45
Riemann-Roch theorem, 30 Simon’s problems, 88
Riemannian geometry, 35 simple, 19
Riemannian manifold, 16 simple group, 55, 56
right angle triangle, 2 simplex algorithm, 43
rigid body, 18 simplicial complex, 11
rigid motion, 13 simplicial spheres, 66
Rising sun lemma, 77 sinc function, 12
Rising sun property, 77 singular continuous measure, 14
Robbins numbers, 79 singular value decomposition, 8
Roessler attractor, 26 singularity theory, 75
Rokhlin lemma, 39 Smale horse shoe, 28, 92
Rokholin tower, 47 Smale’s problems, 88
root, 10 small divisors, 35
root of unity, 37 Smith normal form, 42
rotation, 40 smooth, 18
RSA, 122 smooth function, 10
142
OLIVER KNILL
Smooth Poincare conjecture, 61 symplectic capacity, 68

smooth vector field, 16 symplectic embedding, 68
Sobolev embedding, 48 symplectic form, 16
space group, 40 symplectic manifold, 16, 18
spectral integrability, 39 symplectic matrix, 16
spectral measure, 43 symplectomorphism, 68
Spectral triple, 17 synchronization, 93
spectrum, 15 Szemerédi theorem, 16, 23
sphere, 5, 35
tally stick, 103
Sphere packing, 89
Taxonomy, 92
sphere packing, 41
Taylor series, 12, 28
sphere theorem, 55
tempered distributions, 48
spherical geometry, 35
terminal object, 22
spin gas model, 76
tessarines, 33
splines, 27
tesseract, 17
split algebra, 33
test functions, 48
square matrix, 41
tetrahedron, 17
stability, 36
Thales theorem, 76
standard deviation, 2, 116
Theorem of Hausdorff-Hildebrandt-Schoenberg, 51
standard Dirichlet series, 12
Theorem of Helly, 38
Standard map, 35, 36
Theorem of Lagrange, 31
Standard model, 93
Theorem of Lebowski, 49
standard symplectic form, 16 Theorema egregium, 52
Steinitz theorem, 62 Three circle theorem, 89
Sternberg linarization, 44 three circle theorem, 89
Stirling formula, 86 Thue-Siegel-Roth Theorem, 27
stochastic difference equation, 60 Thurston geometry, 61
stochastic process, 52 tight binding approximation, 93
Stokes, 2 Time-1-map, 44
Stokes theorem, 50 tissue density, 42
Stone-Weierstrass theorem, 27 Toda system, 18
stopping time, 52 Todd class, 37
Strange attractor, 26 Toeplitz matrix, 36
strategy profile, 14 tomography, 93
strong implicit function theorem, 36 topological dynamical system, 23
Strong law of small numbers, 45 topological group, 9
strong Morse inequality, 17 Topological index, 37
structure from motion, 31 topological spaces, 25
Sturm chain, 42 toral dynamical system, 48
Sturm-Liouville theory, 42 toric variety, 67
subgraph, 4 torsion free, 16
subgroup, 21 total curvature, 76
submanifold, 50 transcendental extension, 24
subobject classifier, 22 transcendental numbers, 23
superposition theorem, 34 transformation, 25
supremum norm, 39 translation, 40
surface area of sphere, 29 transport equation, 29
surjective, 3 trigonometric functions, 13
surreal numbers, 103 trigonometry, 12
symbol, 37 trivium, 101
symemtry groyup, 57 Turing computable, 10
symmetric tensor, 16 Turing machine, 10
Symmetry, 90 Tverberg partition, 80
symplectic, 9 Tverberg’s theorem, 80
143
Twin prime, 107 Wedderburn little theorem, 33

Twin prime conjecture, 68 Weierstrass approximation theorem, 26
twist homeomorphism, 61 Weierstrass equation, 57
Tychonov theorem, 5 Weyl conjectures, 90
Whisper galleries, 59
Ulam spiral, 113 Whitney topology, 75
Ullman’s theorem, 31 Wiener problem, 60
ultimate question, 50 Wiener theorem, 27
uncountable, 3 Wigner semi circle law, 44
uniform structure, 78 Wilson primes, 70
uniformization theorem, 59 Wilson theorem, 21, 70, 87
uniformly bounded, 39 winding number, 8
Unimodular map, 39 Witten deformation, 17
Unique prime factorization, 1 Wolf bone, 103
unit ball, 13 Wolfram numbering, 22
unit sphere, 5 Wonderland theorem, 39
unitary group, 9
unitary operator, 6 Yoneda lemma, 10
unitary representation, 10
Zermelo Fraenkel, 72
universal, 41
zero dimensional complexes, 11
Universality, 39
Zeta function, 92
universality, 93
ZFC axiom system, 5
unknot, 32, 76
Zorn lemma, 5
upper bound conjecture, 67
upper bound theorem, 89
valuation, 4, 13
valuation of a field, 88
value, 81
value function, 93
Van der Waerden’s theorem, 46
variance, 19, 116
variational problem, 18
Vector field, 11
vector field, 14
vector space, 9
Vertex Cover, 81
vertex degree, 4
vertices, 4
viral effects, 81
Vitali theorem, 70
Vlasov dynamics, 32
Vlasov system, 32
volume, 15, 29
volume of ball, 29
Wahrscheinlichkeit, 53
wall paper group, 40
Waring problem, 31
Weak convergence, 39
weak Morse inequalities, 17
Weak solution, 33
weak solutions, 18
weak* topology, 64
weakly mixing, 39
144
OLIVER KNILL
Bibliography
References
[1] A005130. The on-line encyclopedia of integer sequences. https://oeis.org.
[2] P. Abad and J. Abad. The hundred greatest theorems.
http://pirate.shu.edu/ kahlnath/Top100.html, 1999.
[3] R. Abraham, J.E. Marsden, and T. Ratiu. Manifolds, Tensor Analysis and Applications. Applied Mathe-
matical Sciences, 75. Springer Verlag, New York etc., second edition, 1988.
[4] A. Aczel. Descartes’s secret notebook, a true tale of Mathematics, Mysticism and the Quest to Understand
the Universe. Broadway Books, 2005.
[5] C.C. Adams. The Knot Book. Freeman and Company, 1994.
[6] D. Adams. The Hitchhiker’s guide to the galaxy. Pan Books, 1979.
[7] R. Aharoni. Mathematics, Poetry and Beauty. World Scientific, 2015.
[8] L. Ahlfors. Complex Analysis. McGraw-Hill Education, 1979.
[9] M. Aigner and G.M. Ziegler. Proofs from the book. Springer Verlag, Berlin, 2 edition, 2010. Chapter 29.
[10] A. Alexander. Duel at Dawn. Harvard University Press, 2010.
[11] P. Alexandroff. Combinatorial topology. Dover books on Mathematics. Dover Publications, Inc, 1960.
Three volumes bound as one.
[12] J.M. Almira and A. Romero. Yet another application of the Gauss-Bonnet Theorem for the sphere. Bull.
Belg. Math. Soc., 14:341–342, 2007.
[13] S. Alpern and V.S. Prasad. Combinatorial proofs of the Conley-Zehnder-Franks theorem on a fixed point
for torus homeomorphisms. Advances in Mathematics, 99:238–247, 1993.
[14] C. Alsina and R.B. Nelsen. Charming Proofs. A journey into Elegant Mathematics, volume 42 of Dolciani
Mathematical Expositions. MAA, 2010.
[15] J.W. Anderson. Hyperbolic Geometry. Springer, 2 edition, 2005.
[16] G.E. Andrews. The theory of partitions. Cambridge Mathematical Library. Cambridge University Press,
1976.
[17] K. Arnold and J. Peyton, editors. A C User’s Guide to ANSI C. Prentice-Hall, Inc., 1992.
[18] V.I. Arnold. Mathematical Methods of Classical Mechanics. Springer Verlag, New York, 2 edition, 1980.
[19] V.I. Arnold. Lectures on Partial Differential Equations. Springer Verlag, 2004.
[20] V.I. Arnold. Experimental Mathematics. AMS, 2015. Translated by Dmitry Fuch and Mark Saul.
[21] E. Asplund and L. Bungart. A first course in integration. Holt, Rinehart and Winston, 1966.
[22] M. Atiyah. K-Theory. W.A. Benjamin, Inc, 1967.
[23] M. Atiyah. Mathematics in the 20th century. American Mathematical Monthly, 108, 2001.
[24] J-P. Aubin and I. Ekeland. Applied nonlinear Analysis. John Wiley and Sons, 1984.
[25] T. Aubin. Nonlinear Analysis on Manifolds. Monge-Amp‘ere equations. Springer, 1982.
[26] J. Bach. The lebowski theorem. https://twitter.com/plinz/status/985249543582355458, Apr 14, 2018.
[27] J.C. Baez. The octonions. Bull. Amer. Math. Soc. (N.S.), 39(2):145–205, 2002.
[28] R. Balakrishnan and K. Ranganathan. A textbook of Graph Theory. Springer, 2012.
[29] W.W. Rouse Ball. A short account of the History of mathematics. McMillan and co, London and New
York, 1988. Reprinted by Dover Publications, 1960.
[30] W. Ballmann. Lectures on Kähler Manifolds. ESI Lectures in Mathematics and Physics.
[31] T. Banchoff. Beyond the Third Dimension, Geometry, Computer Graphics and Higher Dimensions. Sci-
entific American Library, 1990.
[32] A-L. Barabasi. Linked, The New Science of Networks. Perseus Books Group, 2002.
[33] G. Baumslag. Topics in combinatorial group theory. Birkhäuser Verlag, 1993.
[34] A. Beardon. Iteration of Rational Functions. Graduate Texts in Mathematics. Springer-Verlag, New York,
1990.
[35] E. Behrends. Fünf Minuten Mathematik. Vieweg + Teubner, 2006.
[36] E.T. Bell. The Development of Mathematics. McGraw Hill Book Company, 1945.
[37] E.T. Bell. Men of Mathematics. Penguin books, 1953 (originally 1937).
[38] B.Engquist and W. Schmid (Editors). Mathematics Unlimited - 2001 and Beyond. Springer, 2001.
[39] S.K. Berberian. Fundamentals of Real analysis. Springer, 1998.
145
[40] M. Berger. A Panoramic View of Riemannian Geometry. Springer Verlag, Berlin, 2003.
[41] M. Berger and B. Gostiaux. Differential geometry: manifolds, curves, and surfaces, volume 115 of Graduate
Texts in Mathematics. Springer-Verlag, New York, 1988.
[42] W. Berlinghoff and F. Gouvea. Math through the ages. Mathematical Association of America Textbooks,
2004.
[43] H-G. Bigalke. Heinrich Heesch, Kristallgeometrie, Parkettierungen, Vierfarbenforschung. Birkhäuser,
1988.
[44] N.L. Biggs. The roots of combinatorics. Historia Mathematica, 6:109–136, 1979.
[45] G. D. Birkhoff. An extension of poincareés last theorem. Acta Math., 47:297–311, 1925.
[46] J. Bondy and U. Murty. Graph theory, volume 244 of Graduate Texts in Mathematics. Springer, New
York, 2008.
[47] W.W. Boone and G. Higman. An algebraic characterization of the solvability of the word problem. J.
Austral. Math. Soc., 18, 1974.
[48] K.C. Border. Fixed point theorems with applications to economics and game theory. Cambridge University
Press, 1985.
[49] N. Bourbaki. Elements d’histoire des Mathematiques. Springer, 1984.
[50] N. Bourbaki. Elements de Mathematique. Springer, 2006.
[51] J. Bourgain. Green’s function estimates for lattice Schrödinger operators and applications, volume 158 of
Annals of Mathematics Studies. Princeton Univ. Press, Princeton, NJ, 2005.
[52] P. Bourke. Visualising volumetric fractals. GSTF Journal on Computing, 5, 2017.
[53] A. Boutot. Catastrophe theory and its critics. Synthese, 96:167–200, 1993.
[54] C. Boyer. A History of Mathematics. John Wiley and Sons, Inc, 2nd edition, 1991.
[55] F. Brechenmacher. Histoire du theoreme de Jordan de la decomposition matricielle (1870-1930). Ecole des
Hautes Etudes en Sciences Sociales, 2005. PhD thesis, EHESS.
[56] G.E. Bredon. Topology and Geometry, volume 139 of Graduate Texts in Mathematics. Springer Verlag,
1993.
[57] S. Brendle. Ricci Flow and the Sphere theorem. Graduate Studies in Mathematics. AMS, 2010.
[58] D. Bressoud. Historical reflections on the fundamental theorem of calculus. MAA Talk of May 11, 2011.
[59] D. M. Bressoud. Factorization and Primality Testing. Springer Verlag, 1989.
[60] M. Bressoud. Proofs and Confirmations, The story of the Alternating Sign Matrix Conjecture. MMA and
Cambridge University Press, 1999.
[61] O. Bretscher. Calculus I. Lecture Notes Harvard, 2006.
[62] H. Brezis. Functional Anslysis, Sobolev Spaces and Partial Differential Equations. University text.
Springer, 2011.
[63] M. Brown and W.D. Neuman. Proof of the Poincaré-Birkhoff fixed point theorem. Michigan Mathematical
Journal, 24:21–31, 1977.
[64] R.A. Brualdi. Introductory Combinatorics. Pearson Prantice Hall, forth edition, 2004.
[65] G. Van Brummelen. Heavenly Mathematics. Princeton University Press, 2013.
[66] C. Bruter. Mathematics and Modern Art, volume 18 of Springer Proceedings in mathematics. Springer,
2012.
[67] P. Bullen. Dictionary of Inequalities. CRC Press, 2 edition, 2015.
[68] E.B. Burger. Exploring the Number Jungle. AMS, 2000.
[69] W. Burnside. Theory of Groups of Finite Order. Cambridge at the University Press, 1897.
[70] W. Byers. How Mathematicians Think. Princeton University Press, 2007.
[71] W.T. Haight II C. Cavagnaro. Classical and Theoretical mathematics. CRC Press, 2001.
[72] M. Capobianco and J.C. Molluzzo. Examples and Counterexamples in Graph Theory. North-Holland, New
York, 1978.
[73] O. Caramello. Theories, Sites, Toposes. Oxford University Press, 2018.
[74] L. Carleson and T.W. Gamelin. Complex Dynamics. Springer-Verlag, New York, 1993.
[75] A-L. Cauchy. Cours d’Analyse. L’Imprimerie Royale, 1821.
[76] C.Clapman and J. Nicholson. Oxford Concise Dictionary of Mathematics. Oxford Paperback Reference.
Oxford University Press, 1990.
[77] A. Lax C.D. Olds and G. Davidoff. The Geometry of Numbers, volume 41 of Anneli Lax New Mathematical
Library. AMS, 2000.
146
OLIVER KNILL
[78] P. E. Ceruzzi. A History of Modern Computing. MIT Press, second edition, 2003.
[79] K. Chandrasekharan. Introduction to Analytic Number Theory, volume 148 of Grundlehren der mathema-
tischen Wissenschaften. Springer, 1968.
[80] G. Chartrand and P. Zhang. Chromatic Graph Theory. CRC Press, 2009.
[81] W.Y.C. Chen and R. P. Stanley. The g-conjecture for spheres. http://www.billchen.org/unpublished/g-
conjecture/g-conjecture-english.pdf, 2008.
[82] N. Chernof and R. Markarian. Chaotic billiards. AMS, 2006.
[83] C. Chevalley. Theory of Lie Groups. Princeton University Press, 1946.
[84] J.R. Choksi and M.G.Nadkarni. Baire category in spaces of measures, unitary operators and transforma-
tions. In Invariant Subspaces and Allied Topics, pages 147–163. Narosa Publ. Co., New Delhi, 1990.
[85] I. Ciufolini and J.A. Wheeler. Gravitation and inertia. Princeton Series in Physics, 1995.
[86] R. Cochrane. The secret Life of equations: the 50 greatest equations and how they work. Octopus Company,
2016.
[87] R. Cochrane. Math Hacks. Cassell Illustrated, 2018.
[88] E.A. Coddington and N. Levinson. Theory of Ordinary Differential Equations. McGraw-Hill, New York,
1955.
[89] P.R. Comwell. Polyhedra. Cambridge University Press, 1997.
[90] A. Connes. Noncommutative geometry. Academic Press, 1994.
[91] K. Conrad. The origin of representation theory. 2010.
[92] J. Conway. Universal quadratic forms and the fifteen theorem. Contemporary Mathematics, 272:23–26,
1999.
[93] J.B. Conway. Functions of One Complex Variable. Springer Verlag, 2. edition, 1978.
[94] J.B. Conway. A course in functional analysis. Springer Verlag, 1990.
[95] J.B. Conway. Mathematical Connections: A Capstone Course. American Mathematical Society, 2010.
[96] J.H. Conway. On Numbers and Games. A K Peters, Ltd, 2001.
[97] J.H. Conway and R.K. Guy. The book of numbers. Copernicus, 1996.
[98] J.H. Conway and N.J.A.Sloane. What are all the best sphere packings in low dimensions. Discr. Comp.
Geom., 13:383–403, 1995.
[99] J.H. Conway and N.J.A. Sloane. Sphere packings, Lattices and Groups, volume 290 of A series of Com-
prehensive Studies in Mathematics. Springer Verlag, New York, 2.nd edition edition, 1993.
[100] I.P. Cornfeld, S.V.Fomin, and Ya.G.Sinai. Ergodic Theory, volume 115 of Grundlehren der mathematischen
Wissenschaften in Einzeldarstellungen. Springer Verlag, 1982.
[101] R. Courant and H. Robbins. Was ist Mathematik. Springer, fifth edition, 1941.
[102] T. Crilly. 50 mathematical ideas you really need to know. Quercus, 2007.
[103] D. Cristofaro-Gardiner and M. Hutchings. From one reeb orbit to two. https://arxiv.org/abs/1202.4839,
2014.
[104] K.S. Thorne C.W. Misner and J.A. Wheeler. Gravitation. Freeman, San Francisco, 1973.
[105] H.L. Cycon, R.G.Froese, W.Kirsch, and B.Simon. Schrödinger Operators—with Application to Quantum
Mechanics and Global Geometry. Springer-Verlag, 1987.
[106] D.Downing. Dictionary of Mathematical Terms. Barron’s Educational Series, 1995.
[107] M. de Gosson. The symplectic camel principle and semicalssical mechanics. Journal of Physics A: Math-
ematical and General, 35(32):6825, 2002.
[108] W. de Melo and S. van Strien. One dimensional dynamics, volume 25 of Series of modern surveys in
mathematics. Springer Verlag, 1993.
[109] M. Denker, C. Grillenberger, and K. Sigmund. Ergodic Theory on Compact Spaces. Lecture Notes in
Mathematics 527. Springer, 1976.
[110] E. Denne. Alternating quadrisecants of knots. 2004. Thesis at University of Illinois at Urbana-Champaign.
[111] L.E. Dickson. History of the theory of numbers.Vol. I:Divisibility and primality. Chelsea Publishing Co.,
New York, 1966.
[112] L.E. Dickson. History of the theory of numbers.Vol.II:Diophantine analysis. Chelsea Publishing Co., New
York, 1966.
[113] R. Diestel. Graph theory, volume 173 of Graduate Texts in Mathematics. Springer, 5th edition, 2016.
[114] J. Dieudonné. A Panorama of Pure Mathematics. Academic Press, 1982.
[115] J. Dieudonné. Grundzüge der modernen Analysis. Vieweg, dritte edition, 1985.
147
[116] C. Ding, D. Pei, and A. Salomaa. Chinese Remainder Theorem, Applications in Computing, Coding
Cryptography. World Scientific, 1996.
[117] M.P. do Carmo. Differential Forms and Applications. Springer Verlag, 1994.
[118] A. Dold. Lectures on algebraic topology. Springer, 1980.
[119] S. Donaldson and P. Kronheimer. The topology of four manifolds. Clarendon Press, 1990.
[120] J. Doob. Stochastic processes. Wiley series in probability and mathematical statistics. Wiley, New York,
1953.
[121] R. Douady. Application du théorème des tores invariantes. These 3 ème cycle, Université Paris VII, 1982.
[122] B.A. Dubrovin, A.T.Fomenko, and S.P.Novikov. Modern Geometry-Methods and Applications Part I,II,III.
Graduate Texts in Mathematics. Springer Verlag, New York, 1985.
[123] W. Dunham. Journey through Genius. Wiley Science Editions, 1990.
[124] J.-P. Eckmann, H. Koch, and P. Wittwer. A computer-assisted proof of universality for area-preserving
maps. Memoirs of the AMS, 47:1–122, 1984.
[125] J-P. Eckmann and D. Ruelle. Ergodic theory of chaos and strange attractors. Rev. Mod. Phys., 57:617–656,
1985.
[126] B.G. Sidharth (Editor). A century of Ideas. Springer, 2008.
[127] K. Ito (Editor). Encyclopedic Dictionary of Mathematics (2 volumes). MIT Press, second edition edition,
1993.
[128] N.J. Higham (Editor). The Princeton Companion to Applied Mathematics. Princeton University Press,
2015.
[129] R. Brown (Editor). 30-Second Maths. Ivy Press, 2012.
[130] S.G. Krantz (editor). Comprehensive Dictionary of Mathematics. CRC Press, 2001.
[131] S.G. Krantz (Editor). Dictionary of Algebra, Arithmetic and Trigonometry. CRC Press, 2001.
[132] T. Gowers (Editor). The Princeton Companion to Mathematics. Princeton University Press, 2008.
[133] C.H. Edwards. The historical Development of the Calculus. Springer Verlag, 1979.
[134] R. L. Eisenman. Classroom Notes: Spoof of the Fundamental Theorem of Calculus. Amer. Math. Monthly,
68(4):371, 1961.
[135] R. Elwes. Math in 100 Key Breakthroughs. Quercus, 2013.
[136] M. Erickson. Beautiful Mathematics. MAA, 2011.
[137] T. Rokicki et al. God’s number is 20. http://www.cube20.org/, 2010.
[138] R.L. Eubank. A Kalman Filter Primer. Chapman and Hall, CRC, 2006.
[139] H. Eves. Great moments in mathematics (I and II. The Dolciani Mathematical Expositions. Mathematical
Association of America, Washington, D.C., 1981.
[140] K.J. Falconer. Fractal Geometry. Wiley, second edition edition, 2003.
[141] B. Farb and R.K. Dennis. Noncommutative Algebra. Springer, 1993.
[142] B. Farb and J. Wolfson. Resolvent degree, hilbert’s 13th problem and geometry.
https://www.math.uchicago.edu/ farb/papers/RD.pdf, 2018.
[143] O. Faugeras. Three-dimensional computer vision: a geometric viewpoint. Cambridge MA: MIT Press,
Cambridge, MA, USA, second edition edition, 1996.
[144] H. Federer. Geometric measure theory. Die Grundlehren der mathematischen Wissenschaften, Band 153.
Springer-Verlag New York Inc., New York, 1969.
[145] E.A. Fellmann. Leonhard Euler. Birkhäuser, Basel, Boston, Berlin, 2007.
[146] B. Fine and G. Rosenberger. The Fundamental Theorem of Algebra. Undergraduate Texts in Mathematics.
Springer, 1997.
[147] K. Fink. A brief History of Mathematics. OPen Court Publishing Co, 1900.
[148] G. Fischer. Mathematical Models. Springer Spektrum, 2 edition, 2017.
[149] A. Fomenko. Visual Geometry and Topology. Springer-Verlag, Berlin, 1994. From the Russian by Marianna
V. Tsaplina.
[150] J-P. Francoise, G.L. Naber, and T.S. Tsun. Encylopedia of Mathematical Physics. Elsevier, 2006.
[151] T. Franzen. Gödel’s Theorem. A.K. Peters, 2005.
[152] N.A. Friedman. Introduction to Ergodic Theory. Van Nostrand-Reinhold, Princeton, New York, 1970.
[153] R. Fritsch and G. Fritsch. The four-color theorem. Springer-Verlag, New York, 1998. History, topological
foundations, and idea of proof, Translated from the 1994 German original by Julie Peschke.
148
OLIVER KNILL
[154] H. Furstenberg. Recurrence in ergodic theory and combinatorial number theory. Princeton University Press,
Princeton, N.J., 1981. M. B. Porter Lectures.
[155] L. Gamwell. Mathematics + Art, a cultural history. Princeton University Press, 2016.
[156] T.A. Garrity. All the Mathematics you Missed. Cambridge University Press, 2002.
[157] H. Geiges. An Introduction to Contact Geometry, volume 109 of Cambridge studies in advanced mathe-
matics. Cambridge University Press, 2005.
[158] B.R. Gelbaum and J.M.H. Olmsted. Theorems and Counterexamples in Mathematics. Springer, 1990.
[159] J.E. Littlewood G.H. Hardy and G. Polya. Inequalities. Cambridge at the University Press, 1959.
[160] M. Giaquinta and S. Hildebrandt. Calculus of variations. I,II, volume 310 of Grundlehren der Mathema-
tischen Wissenschaften. Springer-Verlag, Berlin, 1996.
[161] E. Girondo and G. González-Diez. Introduction to compact Riemann surfaces and dessins d’enfants, vol-
ume 79. Cambridge University Press, 2012.
[162] P. Glendinning. Math In Minutes. Quercus, 2012.
[163] D. Goldfeld. Beyond the last theorem. The Sciences, 3/4:43–40, 1996.
[164] C. Golé. Symplectic Twist Maps, Global Variational Techniques. World Scientific, 2001.
[165] S.W. Golomb. Rubik’s cube and quarks: Twists on the eight corner cells of rubik’s cube provide a model
for many aspects of quark behavior. American Scientist, 70:257–259, 1982.
[166] C. Gordon, D.L. Webb, and S. Wolpert. One cannot hear the shape of a drum. Bulletin (New Series) of
the American Mathematical Society, 27:134–137, 1992.
[167] T. Gowers. Mathematics, a very short introduction. Oxford University Press, 2002.
[168] R. Graham. Rudiments of Ramsey Theory, volume 45 of Regional conference series in Mathematics. AMS,
1980.
[169] R. Graham. Some of my favorite problems in ramsey theory. Integers: electronic journal of combinatorial
number theory, 7, 2007.
[170] I. Grattan-Guinness. The Rainbow of Mathematics. W. W. Norton, Company, 2000.
[171] B. Green and T. Tao. The primes contain arbitrarily long arithmetic progressions. Annals of Mathematics,
167:481–547, 2008.
[172] P. Griffiths and J. Harris. Principles of Algebraic Geometry. Pure and Applied Mathematics. John Wiley
and Sons, 1978.
[173] H. Groemer. Existenzsätze für Lagerungen im Euklidischen Raum. Mathematische Zeitschrift, 81:260–278,
1963.
[174] J. Gross and J. Yellen, editors. Handbook of graph theory. Discrete Mathematics and its Applications
(Boca Raton). CRC Press, Boca Raton, FL, 2004.
[175] B. Grünbaum. Are your polyhedra the same as my polyhedra? In Discrete and computational geometry,
volume 25 of Algorithms Combin., pages 461–488. Springer, Berlin, 2003.
[176] B. Grünbaum. Convex Polytopes. Springer, 2003.
[177] B. Grünbaum and G.C. Shephard. Tilings and Patterns. Dover Publications, 2013.
[178] R. Guy. The strong law of small numbers. Amer. Math. Monthly, 95:697–712, 1988.
[179] T.C. Hales. Jordan’s proof of the jordan curve theorem. Studies in logic, grammar and rhetorik, 10, 2007.
[180] G.R. Hall. Some examples of permutations modelling area preserving monotone twist maps. Physica D,
28:393–400, 1987.
[181] P. Halmos. Lectures on ergodic theory. The mathematical society of Japan, 1956.
[182] P.R. Halmos. Naive set theory. Van Nostrand Reinhold Company, 1960.
[183] D. Hanson. On a theorem of Sylvester and Schur. Canad. Math. Bull., 16, 1973.
[184] G.H. Hardy. Divergent Series. AMS Chelsea Publishing, 1991.
[185] G.H. Hardy. A mathematician’s Apology. Cambridge University Press, 1994.
[186] G.H. Hardy and M. Riesz. The general theory of Dirichlet’s series. Hafner Publishing Company, 1972.
[187] G.H. Hardy and E.M. Wright. An Introduction to the Theory of Numbers. Oxford University Press, 1980.
[188] R. Hartley and A. Zissermann. Multiple View Geometry in computer Vision. Cambridge UK: Cambridge
University Press, 2003. Second edition.
[189] R. Hartshorne. Algebraic Geometry. Springer-Verlag, New York, 1977. Graduate Texts in Mathematics,
No. 52.
[190] A. Hatcher. Algebraic Topology. Cambridge University Press, 2002.
149
[191] F. Hausdorff. Summationsmethoden und Momentfolgen, I,II. Mathematische Zeitschrift, 9:74–109, 280–
299, 1921.
[192] T. Hawkins. The Mathematics of Frobenius in Context. Sources and Studies in the History of Mathematics
and Physical Sciences. Springer, 2013.
[193] G.A. Hedlund. Endomorphisms and automorphisms of the shift dynamical system. Math. Syst. Theor.,
3:320–375, 1969.
[194] S. Helgason. The Radon transform, volume 5 of Progress in mathematics. Birkhaeuser, 1980.
[195] J.M. Henshaw. An equation for every occasion = 52 formulas + why they matter. Johns Hopkins University
Press, 2014.
[196] M. Herman. Sur la conjugaison différentiable des difféomorphismes du cercle à des rotations. IHES, 49:5–
233, 1979.
[197] R. Hersh. What is Mathematics, Really? Oxford University Press, 1997.
[198] T.H. Hildebrandt and I.J. Schoenberg. On linear functional operations and the moment problem for a
finite interval in one or several dimensions. Annals of Mathematics, 34:317–328, 1933.
[199] H. Hofer and E. Zehnder. Symplectic invariants and Hamiltonian dynamics. Birkhäuser advanced texts.
Birkhäuser Verlag, Basel, 1994.
[200] H. Hofer and E. Zehnder. Symplectic invariants and Hamiltonian Dynamics. Birkhäuser Advanced texts.
Birkhäuser, 1994.
[201] A.J. Hoffman and C.W. Wu. A simple proof of a generalized cauchy-binet theorem. American Mathematical
Monthly, 123:928–930, 2016.
[202] J.B. Hogendijk. Al-mutaman ibn hud, 11the century king of saragossa and brilliant mathematician. His-
toria Mathematica, 22:1–18.
[203] H. Hopf. Über die Drehung der Tangenten und Sehnen ebener Curven. Compositio Math., 2, pages 50–62,
1935.
[204] L.K. Hua. Introduction to Number theory. Springer Verlag, Berlin, 1982.
[205] J. Humphreys. Introduction to Lie Algebras and Representation Theory. Springer, third printing, revised
edition, 1972.
[206] W. Hurewicz. Homotopy and homology. In Proceedings of the International Congress of Mathematicians,
Cambridge, Mass., 1950, vol. 2, pages 344–349. Amer. Math. Soc., Providence, R. I., 1952.
[207] M. Hutchings. Taubes proof of the weinstein conjecture in dimension three. Bull. Amer. Math. Soc.,
47:73–125, 2010.
[208] M. Hutchings. Fun with symplectic embeddings. Slides from Frankferst, Feb 6, 2016.
[209] M. Hutchings and C.H. Taubes. The Weinstein conjecture for stable Hamiltonian structures. Geom. Topol.,
13(2):901–941, 2009.
[210] P. V. M. Blagojević I. Bárány and G.M. Ziegler. Tverberg’s theorem at 50: Extensions and counterexam-
ples. Notices of the AMS, pages 732–739, 2016.
[211] H.S. Zuckerman I. Niven and H.L. Montogmery. An introduction to the theory of numbers. 1991.
[212] O. Lanford III. A computer-assisted proof of the feigenbaum conjectures. Bull. Amer. Math. Soc., 6:427–
434, 1982.
[213] O.E. Lanford III. A shorter proof of the existence of the Feigenbaum fixed point. Commun. Math. Phys,
96:521–538, 1984.
[214] L. Illusie. What is a topos? Notices of the AMS, 51, 2004.
[215] I.R.Shafarevich and A.O. Remizov. Linear Algebra and Geometry. Springer, 2009.
[216] R. Girgensohn J. Borwein, D.Bailey. Experimentation in Mathematics. A.K. Peters, 2004. Computational
Paths to Discovery.
[217] P.C. Jackson. Introduction to artificial intelligence. Dover publications, 1985. Second, Enlarged Edition.
[218] T. Jackson. An illustrated History of Numbers. Shelter Harbor Press, 2012.
[219] T. Jech. The axiom of choice. Dover, 2008.
[220] M. Jeng and O. Knill. Billiards in the lp unit balls of the plane. Chaos, Fractals, Solitons, 7:543–545,
1996.
[221] C. Goodman-Strauss J.H. Conway, H.Burgiel. The Symmetries of Things. A.K. Peterse, Ltd., 2008.
[222] M.C. Jordan. Cours d’Analyse, volume Tome Troisieme. Gauthier-Villards,Imprimeur-Libraire, 1887.
[223] J. Jost. Riemannian Geometry and Geometric Analysis. Springer Verlag, 2005.
150
OLIVER KNILL
[224] D. Joyner. Adventures in Group Theory, Rubik’s Cube, Merlin’s Machine and Other Mathematical Toys.
Johns Hopkins University Press, second edition, 2008.
[225] M. Kac. Can one hear the shape of a drum? Amer. Math. Monthly, 73:1–23, 1966.
[226] R.E. Kalman. A new approach to linear filtering and prediction problems. Transactions of the ASME–
Journal of Basic Engineering, 82(Series D):35–45, 1960.
[227] L.A. Kaluzhnin. The Fundamental Theorem of Arithmetic. Little Mathematics Library. Mir Publishers,
Moscow, 1979.
[228] M. Kaoubi. K-Theory. Springer Verlag, 1978.
[229] A. Katok. Fifty years of entropy in dynamics: 1958-2007. Journal of Modern Dynamics, 1:545–596, 2007.
[230] A. Katok and B. Hasselblatt. Introduction to the modern theory of dynamical systems, volume 54 of
Encyclopedia of Mathematics and its applications. Cambridge University Press, 1995.
[231] A.B. Katok and A.M. Stepin. Approximations in ergodic theory. Russ. Math. Surveys, 22:77–102, 1967.
[232] A.B. Katok and A.M. Stepin. Metric properties of measure preserving homemorphisms. Russ. Math.
Surveys, 25:191–220, 1970.
[233] V. Katz. The history of Stokes theorem. Mathematics Magazine, 52, 1979.
[234] V. Katz. Mathematics of Egypt, Mesopotamia, China, India and Islam. Princeton University Press, 2007.
[235] V.J. Katz. The history of stokes theorem. Mathematics Magazine, 52:146–156, 1979.
[236] Y. Katznelson. An introduction to harmonic analysis. John Wiley and Sons, Inc, New York, 1968.
[237] A.S. Kechris. Classical Descriptive Set Theory, volume 156 of Graduate Texts in Mathematics. Springer-
Verlag, Berlin, 1994.
[238] V.P. Khavin and N.K. Nikol’skij. Commutative Harmonic Analysis I. Springer, 1991.
[239] D.A. Klain and G-C. Rota. Introduction to geometric probability. Lezioni Lincee. Accademia nazionale dei
lincei, 1997.
[240] F. Klein. Vorlesungen über die Entwicklung der Mathematik im 19. Jahrhundert. Springer, 1979 (originally
1926).
[241] J.R. Kline. What is the Jordan curve theorem? American Mathematical Monthly, 49:281–286, 1942.
[242] M. Kline. Mathematical Thought from Ancient to Modern Time. Oxford University Press, 1972.
[243] S. Klymchuk. Counterexamples in Calculus. MAA, 2010.
[244] O. Knill. From the sphere to the mandelbulb. talk May 18, 2013, http://www.math.harvard.edu/ knill/s-
lides/boston/sic.pdf.
[245] O. Knill. Singular continuous spectrum and quantitative rates of weakly mixing. Discrete and continuous
dynamical systems, 4:33–42, 1998.
[246] O. Knill. Weakly mixing invariant tori of hamiltonian systems. Commun. Math. Phys., 205:85–88, 1999.
[247] O. Knill. A multivariable chinese remainder theorem and diophantine approximation.
http://people.brandeis.edu/ kleinboc/EP0405/knill.html, 2005.
[248] O. Knill. Probability Theory and Stochastic Processes with Applications. Overseas Press, 2009.
[249] O. Knill. A discrete Gauss-Bonnet type theorem. Elemente der Mathematik, 67:1–17, 2012.
[250] O. Knill. A multivariable chinese remainder theorem.
https://arxiv.org/abs/1206.5114, 2012.
[251] O. Knill. A Cauchy-Binet theorem for Pseudo determinants. Linear Algebra and its Applications, 459:522–
547, 2014.
[252] O. Knill. Some fundamental theorems in mathematics.
[253] O. Knill. Top 10 fundamental theorems.
https://www.youtube.com/watch?v=Dvj0gfajdNI, 2018.
[254] O. Knill, J. Carlsson, A. Chi, and M. Lezama. An artificial intelligence experiment in college math
education. http://www.math.harvard.edu/˜knill/preprints/sofia.pdf, 2003.
[255] O. Knill and O. Ramirez-Herran. On ullman’s theorem in computer vision. Retrieved February 2, 2009,
from http://arxiv.org/abs/0708.2442, 2007.
[256] O. Knill and E. Slavkovsky. Visualizing mathematics using 3d printers. In C. Fonda E. Canessa and
M. Zennaro, editors, Low-Cost 3D Printing for science, education and Sustainable Development. ICTP,
2013. ISBN-92-95003-48-9.
[257] D. E. Knuth. Surreal numbers. Addison-Wesley Publishing Co., Reading, Mass.-London-Amsterdam, 1974.
[258] T.W. Koerner. Fourier Analysis. Cambridge University Press, 1988.
151
[259] N. Kolmogorov. Grundbegriffe der Wahrscheinlichkeitsrechnung. Berlin, 1933. English: Foundations of

Probability Theory, Chelsea, New York, 1950., 1933.
[260] N. Kondratieva. Three most beautiful mathematical formulas.
http://fy.chalmers.se/˜tfkhj/BeautifulFormulas.pdf, 2006.
[261] J. Kottke. The lebowski theorem of machine superintelligence. https://kottke.org/18/04/the-lebowski-
theorem-of-machine-superintelligence, Apr 16, 2018.
[262] S. Krantz. Function Theory of Serveral Complex Variables. AMS Chelsea publishing, 2 edition, 2001.
[263] S.G. Krantz. Mathematical Apocrypha. AMS, 2002.
[264] S.G. Krantz. Mathematical Apocrypha Redux. AMS, 2005.
[265] S.G. Krantz and H. Parks. A mathematical Odyssey: journal from the real to the complex. Springer Verlag,
2014.
[266] S.N. Krivoshapko and N.N. Ivanov. Encyclopedia of Analytical Surfaces. Springer, 2015.
[267] R. Krömer. Tool and Object: A history and Philosophy of Category Theory. Birkhäuser, 2007.
[268] T. Kuhn. The Structure of Scientific Revolutions, volume 2 of International Encyclopedia of Unified
Science. second edition, enlarged edition, 1970.
[269] G. Kuperberg. Another proof of the alternating-sign matrix conjecture. Internat. Math. Res. Notices,
pages 139–150, 1996.
[270] M. Lackenby. The whitney trick. Topology and its Applications, 71(2):115 – 118, 1996.
[271] I. Lakatos. Proofs and Refutations. Cambridge University Press, 1976.
[272] S. Mac Lane. Categories for the Working Mathematician. Springer, 1998.
[273] O. Lanford. Lecture notes in dynamical systems. ETH Zuerich, 1987.
[274] A.N. Langville and C.D. Meyer. Google’s page rank and beyond: the science of search engine rankings.
Princeton University Press, 2006.
[275] P.D. Lax. Approximation of measure preserving transformations. Commun. Pure Appl. Math., 24:133–135,
1971.
[276] T. Leinster. Rethinking set theory.
[277] S. Lem. The futurological congress. A Harvest book, 1985.
[278] H.W. Lenstra. Solving the pell equation. Algorithmic Number Theory, 44, 2008.
[279] E.H. Lieb and M. Loss. Analysis, volume 14 of Graduate Studies in Mathematics. American Mathematical
Society, 1996.
[280] M. Livio. The equation that couldn’t be solved. Simon and Schuster, 2005.
[281] M. Livio. Is God a mathematician. Simon and Schuster Paperbacks, 2009.
[282] S. Morrison M. Feedman, R. Gompf and K. Walker. Man and machine thinking about the smooth 4-
dimensional poincaré conjecture. https://arxiv.org/pdf/0906.5177.pdf, 2009.
[283] J. MacCormick. 9 Algorithms that changed the future. Princeton University Press, 2012.
[284] B.B. Mandelbrot. The Fractal Geometry of Nature. W.H. Freeman and Company, 1982.
[285] E. Maor. The Pythagorean Theorem: A 4000 year history. Princeton University Press, 2007.
[286] R. Marchthaler and S. Dingler. Kalman-Filter. Springer Verlag, 2017.
[287] P. Maritz and S. Mouton. Francis Guthrie: A colourful life. Mathematical Intelligencer, 34, 2012.
[288] R. Martinez. Proofs that every mathematician should know.
https://math.stackexchange.com/questions/178940/proofs-that-every-mathematician-should-know, 2018.
[289] J-C. Martzloff. A history of Chinese Mathematics. Springer Verlag, second edition, 1997.
[290] B. Mazur. Imagining Numbers. Penguin Books, 2003.
[291] J. Mazur. Fluke, the math and myth of coincidence. Basic Books, 2016.
[292] D. McDuff. What is sympectic geometry. Talk of March 31, 2009, 2009.
[293] D. McDuff and D. Salamon. Introduction to Symplectic Topology. Clarendon Press, 1998.
[294] C. McLarty. The uses and abuses of the history of topos theory. Brit. J. Phil. Sci, pages 351–375, 1990.
[295] C. McLarty. Elementary Categories, Elementary Toposes. Clarendon Press, 1992.
[296] C. McLarty. The rising sea: Grothendieck on simplicity and generality. In Episodes in the history of
modern algebra (1800–1950), volume 32 of Hist. Math., pages 301–325. Amer. Math. Soc., Providence,
RI, 2007.
[297] C. McMullen. Riemann surfaces, dynamics and geometry. Course Notes, Harvard University, 2014.
[298] R. Merris. Graph theory. Interscience Series in discrete mathematics and optimization. Wiley, 2001.
152
OLIVER KNILL
[299] P.W. Michor. Elementary Catastrophe Theory. Tipografia Universitatii din Timisorara, 1985.
[300] J. Milnor. Dynamics in one complex variable. Introductory Lectures, SUNY, 1991.
[301] H. Minc. Nonnegative Matrices. John Wiley and Sons, 1988.
[302] M. Minsky. The Society of Mind. A Touchstone Book, published by Simon @ Shuster Inc, New York,
London, Toronto, Sydney and Tokyo, 1988.
[303] C. Misbah. Complex Dynamics and Morphogenesis and Morphogenesis, An Introduction to Nonlinear
Science. Springer, 2017.
[304] U. Montano. Explaining Beauty in Mathematics: An Aesthetic Theory of Mathematics, volume 370 of
Synthese Library. Springer, 2013.
[305] J.W. Morgan. The poincaré conjecture. In Proceedings of the ICM, 2006, pages 713–736, 2007.
[306] J.W. Morgan and G. Tian. Ricci Flow and the Poincaré Conjecture, volume 3 of Clay Mathematics
Monographs. AMS, Clay Math Institute, 2007.
[307] J. Moser. Stable and random Motion in dynamical systems. Princeton University Press, Princeton, 1973.
[308] J. Moser. Selected chapters in the calculus of variations. Lectures in Mathematics ETH Zürich. Birkhäuser
Verlag, Basel, 2003. Lecture notes by Oliver Knill.
[309] D. Mumford. O. zariski. National Academy of Sciences, 2013.
[310] K.G. Murty. Linear Programming. Wiley, 1983.
[311] M.Ram Murty and A Pacelli. Quadratic reciprocity via theta functions. Proc. Int. Conf. Number Theory,
1:107–116, 2004.
[312] E. Nagel and J.R. Newman. Godel’s proof. New York University Press, 2001.
[313] I.P. Natanson. Constructive Function theory. Frederick Ungar Publishing, 1965.
[314] E. Nelson. Internal set theory: A new approach to nonstandard analysis. Bull. Amer. Math. Soc, 83:1165–
1198, 1977.
[315] J. Neunhäuserer. Schöne Sätze der Mathematik. Springer Spektrum, 2017.
[316] T.Lewis N.I. Fisher and B.J. Embleton. Statistical analysis of spherical data. Cambridge University Press,
1987.
[317] C. Mouhot (notes by T. Feng). Analysis of partial differential equations. 2013. Lectures at Cambridge.
[318] M. Nowak. Evolutionary Dynamics: Exploring the Equations of Life. Harvard University Press, 2006.
[319] O. Ore. The Four-Color Problem. Academic Press, 1967.
[320] J. Owen. The grammar of Ornament. Quaritch, 1910.
[321] J.C. Oxtoby. Measure and Category. Springer Verlag, New York, 1971.
[322] R.S. Palais. Seminar on the Atiyah-Singer Index Theorem, volume 57 of Annals of Mathematics Studies.
Princeton University Press, 1965.
[323] R.S. Palais. A simple proof of the banach contraction principle. Journal Fixed point theory and Applica-
tions, 2:221–223, 2007.
[324] R.S. Palais and S. Smale. A generalized morse theory. Bull. Amer. Math. Soc., 70:165–172, 1964.
[325] E. Pariser. The Filter Bubble. Penguin Books, 2011.
[326] H-O. Peitgen and P.H.Richter. The beauty of fractals. Springer-Verlag, New York, 1986.
[327] C. Petzold. The Annotated Turing: A Guided Tour Through Alan Turing’s Historic Paper on Computabil-
ity and the Turing Machine. Wiley Publishing, 2008.
[328] R. R. Phelps. Lectures on Choquet’s theorem. D.Van Nostrand Company, Princeton, New Jersey, 1966.
[329] G. Pick. Geometrisches zur zahlenlehre. Sitzungsberichte des deutschen naturwissenschaftlich-
medicinischen Vereines für Böhmen ‘Lotos‘ in Prag, Band XIX, 1899.
[330] C.A. Pickover. The Math Book. Sterling, New York, 2012.
[331] B.C. Pierce. Basic Category Theory for Computer Scientists. MIT Press, 1991.
[332] Josep Pla i Carrera. The fundamental theorem of algebra before Carl Friedrich Gauss. Publicacions
Matemàtiques, 36(2B):879–911 (1993), 1992.
[333] B. Polster and M. Ross. Math Goes to the Movies. Johns Hopkins University Press, 2012.
[334] A.S. Posamentier. The Pythagorean Theorem: The Story of its Power and Beauty. Prometheus Books,
2010.
[335] T. Poston and I. Stewart. Catastrophe theory and its applications. Pitman, 1978.
[336] J. Propp. What i learned from richard stanley.
153
[337] Q.I. Rahman and G. Schmeisser. Analytic theory of polynomials, volume 26 of London Mathematical
Society Monographs. Oxford University Press, new series edition, 2002.
[338] A.R. Rajwade and A.K. Bhandari. Surprises and Counterexamples in Real Function Theory. Hindustan
Book Agency, 2007.
[339] F. Rannou. Numerical study of discrete area-preserving mappings. Acta Arithm, 31:289–301, 1974.
[340] M. Reed and B. Simon. Methods of modern mathematical physics. Academic Press, Orlando, 1980.
[341] H. Ricardo. Goldbach’s conjecture implies bertrand’s postulate. Amer. Math. Monthly, 112:492–492, 2005.
[342] D.S. Richeson. Euler’s Gem. Princeton University Press, Princeton, NJ, 2008. The polyhedron formula
and the birth of topology.
[343] H. Riesel. Prime numbers and computer methods for factorization, volume 57 of Progress in Mathematics.
Birkhäuser Boston Inc., 1985.
[344] David P. Robbins. The story of 1,2,7,42,429,7436,... Mathematical Intelligencer, 13:12–19, 1991.
[345] A. Robert. Analyse non standard. Presses polytechniques romandes, 1985.
[346] J. Rognes. On the Atiyah-Singer index theorem. http://www.abelprize.no, 2004.
[347] J. Rosenhouse. The Monty Hall Problem: The Remarkable Story of Math’s Most Contentious Brain Teaser.
Oxford University Press, 2009.
[348] G-C. Rota. The phenomenology of mathematical beauty. Synthese, 111(2):171–182, 1997. Proof and
progress in mathematics (Boston, MA, 1996).
[349] C. Rovelli. The order of Time. Riverhead Books, 2018.
[350] D. Ruelle. Statistical Mechanics. Mathematics Physics Monograph Series. W.A. Benjamin, Inc, 1968.
[351] D. Ruelle. Thermodynamic Formalism. The Mathematical Structures of Classical Equilibrium Statistical
Mechanics, volume 5 of Encyclopedia of mathematics and its applications. Addison-Wesley Publishing
Company, London, 1978.
[352] D. Ruelle. The mathematician’s brain. Princeton University Press, Princeton, NJ, 2007.
[353] J. W. Russell. An Elementary Treatise on Pure Geometry: With Numerous Examples. Clarendon Press,
1893.
[354] S. Russell and P. Norvig. Artificial Intelligence: A modern approach. Prentice Hall, 1995, 2003. second
edition.
[355] D.M.T.Benincasa S. Zeiki, J.P. Romaya and M.F.Atiyah. The experience of mathematical beauty and its
neural correlates. Front. Hum. Neurosci, 13, 2014.
[356] T.L. Saaty and P.C. Kainen. The four color problem, assaults and conquest. Dover Publications, 1986.
[357] A. Sard. The measure of the critical values of differentiable maps. Bull. Amer. Math. Soc., 48:883–890,
1942.
[358] H. Schenk. Computationsl algebraic Geometry, volume 58 of Student texts. Cambridge University Press,
2003.
[359] L. Schläfli. Theorie der Vielfachen Kontinuität. Cornell University Library Digital Collections, 1901.
[360] S.L. Segal. Mathematicians under the Nazis. Princeton university press, 2003.
[361] H. Segerman. Visualizing Mathematics with 3D Printing. John Hopkins University Press, 2016.
[362] G. Shafer and V. Vovk. The sources of komogorov’s grundbegriffe. Statistical Science, pages 70–98, 2006.
[363] C.E. Shannon. A mathematical theory of communication. The Bell System Technical Journal, 27:379–
423,623–656, 1948.
[364] S. Shelah. Proper Forcing. Lecture Notes in Mathematics. Springer, 1982.
[365] T. Siegfried. A beautiful math. Joseph Henry Press, 2006.
[366] J.H. Silverman. The Arithmetic of Elliptic Curves, volume 106 of Graduate Texts in Mathematics. Springer,
1986.
[367] B. Simon. Fifteen problems in mathematical physics. In Anniversary of Oberwolfach, Perspectives in
Mathematics. Birkhäuser Verlag, Basel, 1984.
[368] B. Simon. The statistical mechanics of lattice gases, volume Volume I. Princeton University Press, 1993.
[369] B. Simon. Operators with singular continuous spectrum: I. General operators. Annals of Mathematics,
141:131–145, 1995.
[370] B. Simon. Trace Ideals and their Applications. AMS, 2. edition, 2010.
[371] B. Simon. A comprehensive course in Analysis. AMS, 2017.
[372] Ya.G. Sinai. Introduction to ergodic theory. PrincetonUniversity press, Princeton, 1976.
[373] N. Sinclair. Mathematics and Beauty. Teachers College Press, 2006.
154
OLIVER KNILL
[374] S. Skinner. Sacred Geometry. Sterling, 2006.

[375] J.L. Snell and R. Vanderbei. Three bewitching paradoxes. In Topics in Contemporary Probability and its
Applications, Probability and Stochastics Series, pages 355–370. CRC Press, Boca Raton, 1995.
[376] A. Soifer. The Mathematical Coloring Book: Mathematics of Coloring and the colorful life of its creators.
Springer Verlag, 2009.
[377] R. Solomon. A brief history of the classification of the finite simple groups. Bulletin (new series) of the
AMS, 38:315–352, 2001.
[378] E. Sondheimer and A. Rogerson. Numbers and infinity. Dover Publications Inc., Mineola, NY, 2006. A
historical account of mathematical concepts, Reprint of the 1981 original.
[379] M. Spivak. A comprehensive Introduction to Differential Geometry I-V. Publish or Perish, Inc, Berkeley,
third edition, 1999.
[380] H. Spohn. Large scale dynamics of interacting particles. Texts and monographs in physics. Springer-Verlag,
New York, 1991.
[381] R. Stanley. Combinatorics and Commutative Algebra. Progress in Math. Birkhaeuser, second edition, 1996.
[382] F. Staudacher. Jost Bürgi. Verlag NZZ, 2016.
[383] L.A. Steen and J.A. Seebach. Counterexamples in Topology. Dover, 1995.
[384] W. Stein. Elementary Number Theory: Primes, Congruences, and Secrets. Undergraduate Texts in Math-
ematics. Springer Verlag, 2009.
[385] S. Sternberg. Local contractions and a theorem of Poincaré. Amer. J. Math., 79:809–824, 1957.
[386] S. Sternberg. Lectures on Differential Geometry. Prentice Hall, 1964.
[387] S. Sternberg. A mathematical Companion to Quantum Mechanics. Dover Publications, 2019.
[388] I. Stewart. In Pursuit of the Unknown: 17 equations that changed the world. Basic Books, 2012.
[389] I. Stewart and D. Tall. Algebraic Number Theory and Fermat’s last Theorem. A.K. Peters, third edition,
2002.
[390] J. Stillwell. Mathematics and its history. Springer, 2010.
[391] J. Stillwell. Elements of Mathematics: From Euclid to Goedel. Princeton University Press, 2016.
[392] J.M. Stoyanov. Counterexamples in Probability. Dover, 3nd edition, 2013.
[393] G. Strang. The fundamental theorem of linear algebra. Amer. Math. Monthly, 100(9):848–855, 1993.
[394] S. H. Strogatz. Sync: The Ermerging Science of Spontaneous Order. Hyperion, 2003.
[395] D.W. Stroock. Probability theory, an analytic view. Cambridge University Press, 1993.
[396] D.J. Struik. A concise History of Mathematics. Dover, 1948.
[397] F.J. Swetz and T.I. Kao. Was Pythagoras Chinese? Pennsylvania State University Press, 1988.
[398] S. Tabachnikov. Billiards. Panoramas et synthèses. Société Mathématique de France, 1995.
[399] J. Tanton. Encyclopedia of Mathematics. Facts on File, Inc, 2005.
[400] T. Tao. Structure and Randomness. AMS, 2008.
[401] T. Tao. Topics in Random Matrix Theory. Graduate Studies in Mathematics. AMS, 2012.
[402] C. Tapp. An den Grenzen des Endlichen: Das Hilbertprogramm im Kontext von Formalismus und
Finitismus. Springer Verlag, 2013.
[403] M.E. Taylor. Partial Differential equations, I,II,III, volume 115,116,117 of Applied Mathematical Sciences.
2011.
[404] R.F. Thom. Mathematical Models of Morphogenesis. Mathematics and Its Applications. Horwood Ltd,
1984.
[405] J. Todd. Constructive Theory of Functions. Birkhäuser Verlag, 1963.
[406] E. Trucco and A. Verri. Introductory techniques for 3-D computer vision. Upper Saddle River NJ: Prentice
Hall, 1998.
[407] H. Tverberg. A generalization of radon’s theorem. Journal of the London Mathematical Society, 41:123–
128, 1966.
[408] S. Ullman. The interpretation of visual motion. Cambridge MA : MIT Press, 1979.
[409] G. Urton. Inka History in Knots. University of Texas Press, 2017.
[410] V.Arnold. Catastrophe theory. Springer, 1981.
[411] M. Viazovsaka. The sphere packing problem in dimension 8. http://arxiv.org/abs/1603.04246, 2016.
[412] C. Villani. Optimal transport, old and new. Springer Verlag, 2008.
[413] I.M. Vinogradov. The Method of Trigonometric Sums in the Theory of Numbers. Dover Publications, 1954.
155
[414] D.V. Treshchev V.V. Kozlov. Billiards, volume 89 of Translations of mathematical monographs. AMS,
1991.
[415] Fulton W. Algebraic Curves. An Introduction to Algebraic Geometry. Addison Wesley, 1969.
[416] F. Warner. Foundations of differentiable manifolds and Lie groups, volume 94 of Graduate texts in math-
ematics. Springer, New York, 1983.
[417] D. J. Watts. Small Worlds. Princeton University Press, 1999.
[418] D. J. Watts. Six Degrees. W. W. Norton and Company, 2003.
[419] J.N. Webb. Game Theory. Springer, 2007.
[420] E.W. Weinstein. CRC Concise Encyclopedia of Mathematics. CRC Press, 1999.
[421] J. Weizenbaum. ELIZA—a computer program for the study of natural language communication between
man and machine. Communications of the ACM, 9:36–45, 1965.
[422] D. Wells. Which is the most beautiful? Mathematical Intelligencer, 10:30–31, 1988.
[423] D. Wells. Are these the most beautiful? The Mathematical Intelligencer, 12:9, 1990.
[424] Q. Westrich. The calabi-yau theorem. University of Wisconsin-Madion, 2014.
[425] W.Fulton and J. Harris. Representation theory. Graduate Texts in Mathematics. Springer, 2004.
[426] H. Whitney. Collected Works. Birkhaueser, 1992.
[427] D. Williams. Probability with Martingales. Cambridge mathematical Texbooks, 1991.
[428] R. Wilson. Four Colors Suffice. Princeton Science Library. Princeton University Press, 2002.
[429] G.L. Wise and E.B.Hall. Counterexamples in Probability and Real Analysis. Oxford University Press, 1993.
[430] M. Wojtkowski. Principles for the disign of billiards with nonvanishing lyapunov exponents. Communica-
tions in Mathematical Physics, 105:391–414, 1986.
[431] N. Wolchover. Proven - ’most important unsolved problem’ in numbers. NBC news, 9/11/2012, 2012.
[432] S. Wolfram. Theory and Applications of Cellular Automata. World Scientific, 1986.
[433] S. Wolfram. A new kind of Science. Wolfram Media, 2002.
[434] H. Wussing. 6000 Jahre Mathematik. Springer, 2009.
[435] F.B. Cannonito W.W. Boone and R.C. Lyndon. Word problems. North-Holland Publishing Company,
1973.
[436] D. Zagier. A one-sentence proof that every prime p = 1 mod 4 is a sum of two squares. Amer. Math.
Monthly, 97:144, 1990.
[437] D. Zeilberger. Proof of the alternating-sign matrix conjecture. Elec. J. Combin., 3(Number 2), 1996.
[438] G.M. Ziegler. Lectures on Polytopes. Springer Verlag, 1995.
[439] J. D. Zund. George David Birkhoff and John von Neumann: A question of priority and the ergodic
theorems, 1931-1932. Historia Mathematica, 29:138–156, 2002.
[440] D. Zwillinger. CRC Standard Mathematical Tables and Formulas. CRC Press, 33 edition, 2017.
Department of Mathematics, Harvard University, Cambridge, MA, 02138
156

Fundamental

Uploaded by

Copyright:

Available Formats

Fundamental

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Fundamental

Uploaded by

Copyright:

Available Formats

SOME FUNDAMENTAL THEOREMS IN MATHEMATICS

Abstract. An expository hitchhikers guide to some theorems in mathematics.

Date: 7/22/2018, last update 6/19/2019.

Theorem: Every f ∈ C[x] of degree n can be factored into n linear factors.

Theorem: (X1 + X2 + · · · + Xn ) → Z in distribution.

Theorem: P[A|B] = P[B|A]P[A]/P[B]

Theorem: For d = 2, χ(G) = v − e + f = 2. For d-spheres, χ(G) = 1 + (−1)d .

12. Algebraic geometry

Theorem: ap = a mod p for every prime p and every integer a.

14. Spectral theorem

Theorem: A is normal if and only if A is unitarily diagonalizable.

In finite dimensions, any unitary U diagonalizing A using B = U ∗ AU contains an orthonormal

15. Number systems

all x, y ∈ X. A group is a monoid in which every element x has an inverse y satisfying

Theorem: For analytic f in G and a circle C ⊂ G, one has f (w) = C f(z−w)

G and a function f which is analytic except in a Rfinite set A of points.

18. Linear algebra

Theorem: dim(kerA) + dim(ranA) = n, dim((ranA)⊥ ) = dim(kerAT ).

19. Differential equations

Theorem: If f is Lipschitz, a unique solution of x0 = f (x), x(0) = x0 exists.

Theorem: An axiom system is neither complete nor provably consistent.

21. Representation theory

22. Lie theory

Theorem: The generally recursive class is the Turing computable class.

24. Category theory

Theorem: N (hA , F ) can be identified with F (A).

25. Perturbation theory

Theorem: χ(X × Y ) = χ(X)χ(Y ).

27. Metric spaces

Theorem: A contraction has a unique fixed point in X.

28. Dirichlet series

Theorem: limx→0 sin(x)/x = 1.

Theorem: log(uv) = log(u) + log(v).

31. Geometric probability

Theorem: The space of valuations is (n + 1)-dimensional.

32. Partial differential equations

33. Game theory

34. Measure theory

Theorem: µ = µac + µsc + µpp .

35. Geometric number theory

Theorem: If K is Minkowski and |K| > 2n |Γ| then K ∩ Γ 6= ∅.

Theorem: Either ∃f 6= 0 with Af = 0 or for all g, ∃f with Af = g.

37. Prime distribution

Theorem: The set of primes contains arbitrary long arithmetic progressions.

38. Riemannian geometry

39. Symplectic geometry

Theorem: Every symplectic form is locally diffeomorphic to ω0 .

40. Differential topology

Theorem: ck − ck−1 + · · · + (−1)k c0 ≥ bk − bk−1 + · · · + (−1)k b0 .

41. Non-commutative geometry

Theorem: δ(x, y) = supf ∈A,||[D,f ]||≤1 |f (x) − f (y)|.

Theorem: There are 5 platonic solids and p = (∞, 5, 6, 3, 3, 3, . . . )

43. Descriptive set theory

Theorem: Complete metric spaces are of second Baire category.

44. Calculus of variations

45. Integrable systems

Theorem: For a Liouville integrable system, level surfaces F = c are tori.

An example how to get integrals is to write the system as an isospectral deformation of

46. Harmonic analysis