A RANDOM WALK IN REPRESENTATIONS

Shanshan Ding

A DISSERTATION

in

Mathematics

Presented to the Faculties of the University of Pennsylvania

in

Partial Fulfillment of the Requirements for the

Degree of Doctor of Philosophy

2014

Robin Pemantle, Merriam Term Professor of Mathematics

Supervisor of Dissertation

David Harbater, C. H. Browne Distinguished Professor of Mathematics

Graduate Group Chairperson

Dissertation Committee:
Jonathan Block, Professor of Mathematics
Robin Pemantle, Merriam Term Professor of Mathematics
Martha Yip, Lecturer of Mathematics
A RANDOM WALK IN REPRESENTATIONS

© COPYRIGHT

2014

Shanshan Ding
Acknowledgments

In the sunset of dissolution, everything is illuminated by the aura of

nostalgia, even the guillotine.

– Milan Kundera, The Unbearable Lightness Of Being

First I must thank my advisor Robin Pemantle for, among many other things,

introducing me to probability at a time when I was lost for a direction, suggesting

the immensely satisfying topic of this dissertation, and keeping me on track while

supporting whatever I wish to pursue. I am also indebted to Martha Yip for her

indispensable technical assistance and general awesomeness; she always has answers

to my questions and is invariably generous with her time. Moreover, I am grateful

for Jim Haglund’s advice and encouragement in the later stages of this work.

The graduate chairs during my time at Penn have all been very sympathetic.

Tony Pantev and David Harbater helped me navigate the beginning and the end of

grad school, while Jonathan Block refused to let me give up on myself through all

the years in the middle.

I would like to thank Persi Diaconis for pioneering a beautiful field of mathemat-

iii
ics at the intersection of probability and representation theory. Just as importantly,

I thank Ehrhard Behrends, whom I have never met, but without whose wonderfully

clear introduction to this field I may still be searching for a thesis topic.

I want to acknowledge the professors who inspired me to go to grad school and

whose recommendation letters made it possible: Robert Friedman, Dorian Goldfeld,

John Morgan, and Ken Ono. In fact, I am deeply grateful for all of the outstanding

teachers and mentors, in and out of math, that I have encountered from grade school

to grad school.

At the same time, I also thank my students, who have collectively taught me

much more than I can ever teach them. Their joy in moments of breakthrough

reminded me of why I chose to study math on those occasions when it was otherwise

difficult to remember.

The department’s administrative staff—Janet Burns, Monica Pallanti, Paula

Scarborough, and Robin Toney—steadfastly protects us and holds the department

together. They bring sense and warmth into a building that I think has been proved

to be topologically baffling and bleak. I wish Janet the very best in her upcoming

retirement, though a DRL without her is an unfathomable place.

I am proud to have been part of the Penn math department’s graduate student

community, which has consistently been comprised of friendly, talented, and all-

around wonderful people. It is impossible to name everyone as the requisite margins

of this thesis are too large, but I would be remiss to not at least reference a subset.

iv
In particular, I would like to thank Ying Zhao for being my officemate during an

unforgettable year. I am also much indebted to Hilaf Hasson for his wisdom and

patience. Michael Lugo almost single-handedly taught me to think like a probabilist.

Deborah Crook, from Montreal to Marrakech, and from California to Cappadocia,

it has been fun traveling half of the world with you, and here’s to hoping that we

get to the other half at some point. Paul Levande, I already miss your insight

and wit, and I would be so very sad to miss watching in person your reaction to

Ben Affleck’s Batman. Additionally, I am grateful for the friendship and extensive

mathematical assistance over the years from Adam Topaz, David Lonoff, Elaine So,

Matthew Wright, Matti Astrand, Ryan and Better Ryan (Eberhart and Manion,

not necessarily in that order), Taisong Jing, Torin Greenwood, Tyler Kelly, Ying

Zhang, and Zhentao Lu.

I would like to thank the Philadelphia Orchestra for affording students in the

area the weekly opportunity to attend world-class concerts. These concerts have

been reminders on many Saturdays that, high or low, light or heavy, all notes

eventually come to pass as a rich, transformative melody is left behind.

Finally, it’s not always sunny in Philadelphia, and I extend my sincerest grat-

itude to all those whose optimism and support, in good times and bad, sustained

me through the storms.

The writing of this dissertation is partially supported by NSF grant DMS-

1209117.

v
 ABSTRACT

A RANDOM WALK IN REPRESENTATIONS

Shanshan Ding

Robin Pemantle

The unifying objective of this thesis is to find the mixing time of the Markov chain

on Sn formed by applying a random n-cycle to a deck of n cards and following

with repeated random transpositions. This process can be viewed as a Markov

chain on the partitions of n that starts at (n), making it a natural counterpart to

the random transposition walk, which starts at (1n ). By considering the Fourier

transform of the increment distribution on the group representations of Sn and then

computing the characters of the representations, Diaconis and Shahshahani showed

in [DS81] that the order of mixing for the random transposition walks is n ln n. We

adapt this approach to find an upper bound for the mixing time of the n-cycle-to-

transpositions shuffle. To obtain a lower bound, we derive the distribution of the

number of fixed points for the chain using the method of moments. In the process,

we give a nice closed-form formula for the irreducible representation decomposition

of tensor powers of the defining representation of Sn . Along the way, we also look at

the more general m-cycle-to-transpositions chain (m ≤ n) and give an upper bound

for the mixing time of the m = n − 1 case as well as characterize the expected

number of fixed points in the general case where m is an arbitrary function of n.

vi
Contents

1 Introduction 1

1.1 A mathematical history of card shuffling . . . . . . . . . . . . . . . 1

1.2 Scope and organization of this thesis . . . . . . . . . . . . . . . . . 5

2 Technical Preparations 9

2.1 Markov chains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2.2 Harmonic analysis on finite groups . . . . . . . . . . . . . . . . . . 12

2.3 Representation theory of Sn . . . . . . . . . . . . . . . . . . . . . . 20

3 Upper Bound 28

3.1 The m = n case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

3.2 The m = n − 1 case . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

4 Fixed Points and Lower Bound 37

4.1 Fixed points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

4.2 The defining representation . . . . . . . . . . . . . . . . . . . . . . 39

vii
 4.3 Lower bound for the m = n case . . . . . . . . . . . . . . . . . . . . 45

5 Further Considerations 50

5.1 Miscellaneous results . . . . . . . . . . . . . . . . . . . . . . . . . . 50

5.2 Open questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

Bibliography 56

viii
List of Figures

2.1 Young diagrams corresponding to the partitions of 4 . . . . . . . . . 21

2.2 H1,2 and the array of hook lengths for (4, 4, 3) . . . . . . . . . . . . 23

2.3 Examples and non-examples of rim hooks . . . . . . . . . . . . . . . 24

(4,4,3)
2.4 Computing χ(5,4,2) with the Murnaghan-Nakayama rule . . . . . . . 25

3.1 Examples of λ for which χλ(n−1,1) 6= 0 . . . . . . . . . . . . . . . . . 32

ix
Chapter 1

Introduction

The goals we pursue are always veiled.

– The Unbearable Lightness Of Being, “Words Misunderstood”

1.1 A mathematical history of card shuffling

We start with a question of centuries-old interest to diviners, gamblers, and magi-

cians: how many shuffles does it take to mix a deck of cards?

Naturally, the answer depends on what we mean by “shuffle” and “mix”. Broadly

speaking, a shuffle on n cards is a permutation of the set {1, 2, . . . n} by an element

σ of the symmetric group Sn . The outcome of a sequence of shuffles σ1 , σ2 , . . . , σk

is then permutation by the composition σk · · · σ2 σ1 . We presume that each σi

is chosen according to some probability distribution on Sn , so that the sequence

σ1 , σ2 σ1 , σ3 σ2 σ1 , . . . forms a Markov chain on Sn . If, furthermore, each σi is chosen

1
from the same distribution, then this chain is a random walk on Sn . The distribu-

tion of σk · · · σ2 σ1 is a probability measure on Sn for each k, and the deck is mixed

when the total variation distance between this measure and the uniform measure

on Sn is small. Intuitively, mixing means that one can no longer infer the positions

of the cards from their initial order.

Since the early 1900s and especially during the past 30 years, mathematical

analyses of card shuffling have inspired significant progress in the theory of Markov

chain mixing times, particularly in revealing its rich connections with algebraic

combinatorics. Markov himself had cited card shuffling as a leading example of his

eponymous processes, and his 1906 proof in [Mar06] for the convergence of finite-

state Markov chains implies that shuffling eventually mixes the deck. Poincaré then

supplied a Fourier-analytic proof in [Poi12].

Of course, eventual mixing has always been the implicit premise of card shuffling,

so the more pertinent question is how soon. The first significant breakthrough

in this topic came in 1981, when Diaconis (a former professional magician) and

Shahshahani showed in [DS81] that the order of mixing for the random transposition

shuffle, where one repeatedly chooses two random cards and exchanges them, is

n ln n. Though this shuffle is unlikely to be employed by card players in real life,

[DS81] is a landmark development in probability theory for introducing techniques

from representation theory. A very high-level summary of its ideas is as follows:

Fourier transforms convert convolutions of probability distributions in the time

2
domain to pointwise products in the frequency domain, and the “frequency domain”

for a non-abelian group is given by its group representations, so we can track the

mixing of a Markov chain by observing the Fourier transform of the increment

distribution on the representations of the underlying group, which in turn can be

quantified by computing and summing the characters of the representations. This

approach is applicable to all random walks generated by conjugacy classes of finite

groups, and it was used by Hildebrand ([Hil92]: random transvections in SLn (Fq )),

Pemantle ([Pem94]: 3-cycles), and Lulov ([Lul96]: a wide class of fixed-point-free

permutations, including fixed-point-free involutions) to obtain the mixing times of

various other random walks that fit the description.

Meanwhile, further techniques arose from studies of Markov chains that more

realistically model human card shuffling. Aldous and Diaconis [AD86] introduced

the concept of strong stationary time to prove that the order of mixing for the top-

to-random shuffle, where the top card is removed and inserted into the deck at a

random position, is n ln n. Using a coupling construction, Pemantle [Pem89] proved

an upper bound of O(n2 ln n) for the overhand shuffle, where one shaves off packets of

cards from the top of the deck and stacks each packet on top of the previous one until

all cards have been transferred to the new pile. This bound was ultimately shown to

be tight by Jonasson [Jon06] using a method for establishing sharp lower bounds due

to Wilson [Wil04]. As for the riffle shuffle, the most common shuffling technique

where one divides a deck into two piles and interlaces them together, Bayer and

3
Diaconis [BD92] concluded that seven shuffles are necessary and sufficient to mix a

52-card deck and in the process related the underlying Markov chain to Solomon’s

descent algebra and Hochschild homology. Throughout the 1990s, extensions of

the techniques used to study card shuffling have led to substantial progress in the

general study of stochastic processes on groups, including diffusions on Lie groups

(see [S-C01]). For comprehensive surveys of the works produced, refer to [Dia01]

and [S-C04].

Research in card shuffling and related topics is active and ongoing. Recent areas

of focus have included systematic scan versions of well-understood shuffles, whereby

the location of each update is deterministic ([MPS04], [MNP12]), and randomization

of only selected features, such as card values but not suits [CV06] or the location

of the original bottom card [ADS11]. Extensive effort has also been devoted to

exploring and leveraging the symbiotic connections between card shuffling and the

theories of Lie type groups ([Ful00], [Ful01]), quasi-symmetric functions ([Sta01],

[DF09]), and Hopf algebras [DPR12]. The pervasive theme in this line of research

since Diaconis and Shahshahani’s analysis of the random transposition shuffle has

been the marrying of spectral and probabilistic phenomena and techniques, a theme

that reverberates in the modern studies of expander graphs (see [HLW06]) and

random matrices (see [Tao12]).

As Markov chains have a wide range of applications, any new development in

the field has built-in ramifications for potentially multiple areas of applied math.

4
Two standout applications derived specifically from the work on card shuffling are

in cryptography ([Mir02], [HMR12]), where shuffles are exploited as enciphering

schemes, and in genetics ([SG89], [Dur03]), where shuffles model rearrangements of

DNA segments. Of course, we should not overlook the implications of card shuffling

research for card playing itself ([Tho73], [CH10]). Vegas certainly paid attention and

even invited Diaconis, the renegade magician, for a homecoming of sorts to assess

some new automated shuffling machines. For the findings of the said investigation,

see [DFH13], though we take this opportunity to put forth the disclaimer that no

knowledge of gambling will be endorsed or imparted, here and throughout.

1.2 Scope and organization of this thesis

Nearly the entirety of the the card shuffling literature that we just surveyed deals

with time-homogeneous Markov chains, where the same method of shuffling is re-

peated until the deck is mixed. The present thesis, on the other hand, is motivated

by a time-inhomogeneous Markov chain: after a single application of an n-cycle to

a deck of n cards, how many transpositions are needed to mix the deck?

This chain is a natural counterpart to the random transposition walk on Sn

in the following sense: a transposition changes the cycle type of a permutation

by either splitting a cycle in two (if the two objects transposed are in the same

cycle) or joining two cycles as one (if the two objects are in different cycles), so

random transpositions in fact induce a Markov chain on the set of partitions of n;

5
the time-homogeneous random transposition walk is one such chain that starts at

the partition (1n ), whereas the process we proposed is one that starts at the other

extreme, (n). Markov chains formed on partitions under random transpositions are

examples of coagulation-fragmentation processes, the profound mathematics and

applications of which are surveyed in [Ald99], and a related chain whose eigenfunc-

tions give probabilistic interpretations for the Macdonald polynomials is constructed

in [DR12].

The focus of this thesis is the n-cycle-to-transpositions chain viewed as a pro-

cess on Sn instead of on the partitions of n, though we do hope that our work can

lead to new insight on coagulation-fragmentation processes. We will in fact con-

sider the more general process of a random m-cycle (m ≤ n) followed by random

transpositions. Formally:

Question. Fix m as a function m(n) of n with 2 ≤ m ≤ n for all n. Form a Markov

chain {Xk } on the symmetric group Sn as follows: let X0 be the identity1 , set

X1 = πX0 , where π is a uniformly selected m-cycle, and for k ≥ 2 set Xk = τk Xk−1 ,

where τk is a uniformly selected transposition. Observe that Xk ∈ An when m and

k are of the same parity. Otherwise, Xk ∈ Sn \An . Let µk be the law of Xk , and let

Uk be the uniform measure on An if Xk ∈ An and the uniform measure on Sn \An

if Xk ∈ Sn \An . What is the total variation distance between µk and Uk ?

The increment distributions of these Markov chains are conjugacy-invariant, so

1
Markov chains on finite groups are translation-invariant, so setting X0 to some other element
of Sn may affect parity, but not mixing time.

6
we follow Diaconis and Shahshahani’s approach and adapt their analysis of the

random transposition shuffle to obtain upper bounds for the mixing times of the

m = n and m = n − 1 cases. The relevant concepts and tools from probability,

Fourier analysis, and representation theory are introduced and modified as necessary

in Chapter 2, while the computations are carried out in Chapter 3. The lower bound

was much trickier, but we ultimately obtain one for the m = n case in Chapter 4

by deriving the distribution of the number of fixed points using the method of

moments. Putting the two together gives the main result:

Theorem 3.1.1 and Corollary 4.3.5. For any c > 0, after one n-cycle and cn

transpositions,

e−2c e−2c
− o(1) ≤ kµcn+1 − Ucn+1 kTV ≤ √ + o(1)
e 2 1 − e−4c

as n goes to infinity.

Our arguably most significant contribution is that, while trying to compute the

moments of the fixed point distribution, we discovered a neat (in all senses of the

word) formula for the decomposition of tensor powers of the defining representation

(see Definition 4.2.1) % of Sn :

Theorem 4.2.3. Let λ = (λ1 , λ2 , . . . , λl ) be a partition of n, and let S λ denote the

irreducible representation of Sn corresponding to the shape λ. For 1 ≤ r ≤ n − λ2 ,

the multiplicity aλ,r of S λ in the irreducible representation decomposition of %⊗r is

7
given by

r   
X
λ̄ i r
aλ,r =f ,
|λ̄| i
i=|λ̄|

where λ̄ is the truncated partition (λ2 , . . . , λl ) of weight |λ̄|, f λ̄ is the number of

r
standard Young tableaux of shape λ̄, and i
is a Stirling number of the second

kind.

In Chapter 5 we give two more results on expected numbers of fixed points,

one about the m-cycle-to-transpositions chain for arbitrary m, and the other the

following little gem:

Proposition 5.1.1. If a Markov chain on Sn whose increment distributions are

class measures starts with one fixed point, then it will always average exactly one

fixed point.

We then conclude by reflecting on what could have been and what could still

be, enumerating questions that seem just out of reach and suggesting related topics

that may be within grasp.

8
Chapter 2

Technical Preparations

Without realizing it, the individual composes his life according to

the laws of beauty even in times of greatest distress.

– The Unbearable Lightness Of Being, “Soul and Body”

2.1 Markov chains

As the King asked of the White Rabbit, we begin at the beginning. Specifically, we

begin with a very brief introduction to the central objects of this thesis: Markov

chains. For comprehensive treatises, check out [LPW08] or [Beh00].

Definition 2.1.1. A sequence of random variables (X0 , X1 , . . .) is a Markov chain

on a finite set Ω if, for all xi ∈ Ω and k ≥ 1,

P(Xk+1 = xk+1 | Xk = xk ) = P(Xk+1 = xk+1 | X0 = x0 , . . . , Xk = xk ). (2.1.1)

In words, given the present, the future is independent of the past.

9
 When a Markov chain is at state x, the next position is chosen according to a

fixed probability distribution P (x, ·). If every step is chosen according to the same

transition matrix P , then the chain is said to be time-homogeneous, and the k-step

transition probabilities are given by P k .

If Ω is a finite group, a probability distribution µ on Ω induces a Markov chain

with transition probabilities P (x, yx) := µ(y). This means that the chain moves via

left multiplication by a random element of Ω selected according to µ. The measure

µ is called the increment distribution on Ω.

Definition 2.1.2. A chain is irreducible if it is possible to get from any state to

any other state.

Definition 2.1.3. Let T (x) be the set of times when it is possible for a chain

starting at state x to return to x. The period of x is the gcd of T (x). The chain is

aperiodic if all states have period one.

Definition 2.1.4. For a time-homogeneous Markov chain on Ω with transition

matrix P , a distribution π on Ω satisfying πP = π is a stationary distribution of

the chain.

If Ω is a finite group, then for a chain on Ω with increment distribution µ, the

uniform distribution UΩ satisfies

X 1 X 1 X 1
UΩ (y)P (y, x) = P (y, x) = µ(xy −1 ) = = UΩ (x) (2.1.2)
y∈Ω
|Ω| y∈Ω |Ω| y∈Ω |Ω|

10
for all x ∈ Ω, as the second to last equality is from the observation that the operation

y → xy −1 re-indexes Ω. Thus the uniform distribution is a stationary distribution

for Markov chains on finite groups. Note that as a result, the distance to stationarity

does not depend on the initial state: a chain that starts at x is simply a translation

by x of a chain starting at the identity element, and the uniform distribution is

translation-invariant.

Most of the theory on finite-state Markov chains is concerned with the long-

term behavior of the chains. In particular, we would like to know whether a chain

converges to a stationary distribution and, if so, how quickly. To quantify the speed

of convergence, we need an appropriate metric for measuring the distance between

probability distributions.

Definition 2.1.5. The total variation distance between measures µ and ν on Ω is

1X
kµ − νkTV = |µ(x) − ν(x)| = max|µ(A) − ν(A)|. (2.1.3)
2 x∈Ω A⊆Ω

Theorem 2.1.6 (Markov chain convergence theorem). Every time-homogeneous,

irreducible, and aperiodic Markov chain has a unique stationary distribution π.

Furthermore, there exist constants 0 < α < 1 and C > 0 such that

max kP k (x, ·) − πkTV ≤ Cαk . (2.1.4)

x∈Ω

Proof. See Theorem 4.9 of [LPW08].

The convergence theorem states the sufficient condition for mixing and even

11
specifies that mixing is exponentially fast. However, it gives no information on how

to determine the actual rate of convergence, which typically needs to be handled

on a case-by-case basis.

Before moving on, we should note that the Markov chain defined in Chapter 1

is time-inhomogeneous and periodic. While periodicity will present complications

in the next section, it is not difficult to see that this chain alternates between

An and Sn \An , and that each of the two subsequences converges to the uniform

distribution on the corresponding coset. It is also clear that inhomogeneity does

not affect whether a chain converges as long as the chain is time-homogeneous after

a finite number of steps.

2.2 Harmonic analysis on finite groups

In this section we present an overview of Diaconis and Shahshahani’s approach to

analyzing Markov chain mixing times. A detailed and accessible treatment of the

material can be found in Chapters 15 and 16 of [Beh00]. Another helpful resource

is Chapter 3 of [CST08].

In what follows, let G be a finite group.

Definition 2.2.1. A d-dimensional (unitary) representation ρ of G is a group ho-

momorphism from G to the set of d-by-d unitary matrices, that is, ρ(gh) = ρ(g)ρ(h)

for all g, h ∈ G. The 1-dimensional representation that sends every g ∈ G to 1 is

the trivial representation ρtriv of G.

12
Example. The 1-dimensional representation of Sn which is 1 on An and −1 on Sn \An

is the sign representation of Sn .

Definition 2.2.2. (1) Representations ρ0 and ρ00 of the same dimension d are equiv-

alent if there exists a d-by-d unitary matrix M such that ρ2 (g) = M ρ1 (g)M −1 for

all g ∈ G.

(2) A representation ρ is irreducible, an irrep for short, if it is not equivalent to a

representation of the form ρ1 ⊕ ρ2 .

Remark. By Maschke’s theorem (see, for instance, Theorem 1.5.3 of [Sag01]), every

representation of a finite group is equivalent to a direct sum of irreps.

An alternative way to characterize representations of G is in terms of the vector

spaces that elements of G act on.

Definition 2.2.3. A vector space V is a G-module if there is a G-action · on V

such that g · (av + bw) = a(g · v) + b(g · w) for all g ∈ G, v, w ∈ V , and a, b ∈ C.

We say that a G-module V carries a representation of G. Two representations

are equivalent if their associated G-modules are isomorphic, and a representation is

irreducible if its associated G-module contains no non-trivial G-submodule.

Remark. To go back and forth between Definitions 2.2.1 and 2.2.3, define the group

action g · v to be (ρ(g))(v).

We use Ĝ to denote a collection2 of representations of G that contains precisely

2
If G is abelian, then all representations of G are 1-dimensional and Ĝ is a group itself, com-
monly referred to as the Pontryagin dual of G.

13
one representative from each equivalence class of irreps of G.

Definition 2.2.4. Let f be a function on G. The Fourier transform of f is the

matrix-valued map on Ĝ defined by fˆ(ρ) = g∈G f (g)ρ(g).

P

The key idea of Diaconis and Shahshahani’s approach is to translate the question

of “how close to uniformity is µ” to “how close to 0 (the zero matrix) is µ̂ on the

non-trivial3 irreps of G”. The following helps to start making this idea precise:

Theorem 2.2.5 (Plancherel’s formula). For any function f on G,

X 1 X
|f (g)|2 = dρ tr[fˆ(ρ)(fˆ(ρ))† ], (2.2.1)
g∈G
|G|
ρ∈Ĝ

where dρ is the dimension of ρ and (fˆ(ρ))† is the conjugate transpose of fˆ(ρ).

Proof. See Proposition 16.16 of [Beh00].

Remark. Theorem 2.2.5 is a consequence of the celebrated Peter-Weyl theorem,

which says that the collection of normalized coordinate functions

q 
ρ
dρ /|G|ϕij : ρ ∈ Ĝ, 1 ≤ i, j ≤ dρ , (2.2.2)

where ϕρij is defined by assigning to ϕρij (g) the ij-th entry of ρ(g), is an orthonormal

basis for the space4 of L2 functions on G. The Peter-Weyl theorem applies to all

compact topological groups; a proof for the case of finite groups is given in Theorem

16.11 of [Beh00].

3
As we will see, the sign representation is also excluded for the m-cycle-to-transpositions chain
due to the parity of the chain.
4
P
This is a Hilbert space with the inner product hf1 , f2 iG = g∈G f1 (g)f2 (g).

14
 Let µ and ν be measures on G. Rewriting Theorem 2.2.5 with f = µ − ν gives

X 1 X
(µ(g) − ν(g))2 = dρ tr[(µ̂(ρ) − ν̂(ρ))(µ̂(ρ) − ν̂(ρ))† ], (2.2.3)
g∈G
|G|
ρ∈Ĝ

and the connection to mixing begins to emerge.

If a Markov chain {X0 , X1 , . . .} on G is time-homogeneous with increment dis-

tribution υ, then the k-step move from X0 to Xk is governed by υ ∗k , the k-fold

convolution of υ.

Definition 2.2.6. Let υ and η be measures on G. Their convolution is the measure

υ(gh−1 )η(h).
P
defined by (υ ∗ η)(g) = h∈G

∗k = υ̂ k .
∗ η = υ̂ η̂. Thus υc
Proposition 2.2.7. For any υ and η on G, υ[

Proof. See Proposition 16.19 of [Beh00].

Proposition 2.2.8. If µ is a symmetric measure, i.e. if µ(g) = µ(g −1 ) for all

g ∈ G, then µ̂(ρ) = (µ̂(ρ))† for all ρ ∈ Ĝ.

Proof. See Lemma 16.23 of [Beh00].

P
Proposition 2.2.9. If ρ is any non-trivial irrep of G, then g∈G ρ(g) = 0, and

hence U
cG (ρ) = 0 for the uniform measure UG on G.

Proof. 5 Since ρ is non-trivial, there exists g0 ∈ G such that ρ(g0 ) 6= Idρ , and

X X X
ρ(g) = ρ(g0 g) = ρ(g0 ) ρ(g). (2.2.4)
g∈G g∈G g∈G

5
Despite being widely used, we have not found a coherent proof of this proposition in any text.
The proof given here is adapted from the proof of Lemma 15.3 of [Beh00], which is for the special
case where G is abelian.

15
Consider V , the G-module that carries the representation ρ. It is straightforward
P
to verify that W = {( g∈G ρ(g))(v) : v ∈ V } is a G-submodule of V . Since ρ is
P
irreducible, W must be either trivial or V itself. If W is trivial, then g∈G ρ(g) = 0,
P P
and if W = V , then g∈G ρ(g) is invertible. But g∈G ρ(g) cannot be invertible
P
because ρ(g0 ) 6= Idρ in (2.2.4), so it must be that g∈G ρ(g) = 0.

Suppose that υ is symmetric. Furthermore, suppose that {Xk } is aperiodic and

irreducible, so that υ ∗k converges to UG . Applying Propositions 2.2.7-2.2.9 and the

observation that µ̂(ρtriv ) = 1 for any µ to (2.2.3) gives the L2 distance

X
2 1 X
υ ∗k (g) − UG (g) = dρ tr[(υ̂(ρ))2k ] (2.2.5)
g∈G
|G|
ρ∈Ĝ
ρ6=ρtriv

between υ ∗k and UG .

For arbitrary x1 , . . . , xj , the Cauchy-Schwarz inequality implies that

j j
X X
2
( xi ) ≤ j x2i , (2.2.6)
i=1 i=1

which, applied to Definition 2.1.5, gives that

!2
X X
4kµ − νk2TV = |µ(g) − ν(g)| ≤ |G| (µ(g) − ν(g))2 . (2.2.7)
g∈G g∈G

This extracts from (2.2.5) the upper bound

X
4kυ ∗k − UG k2TV ≤ dρ tr[(υ̂(ρ))2k ] (2.2.8)
ρ∈Ĝ
ρ6=ρtriv

for the total variation distance between υ ∗k and UG .

16
 Before we work on the right hand sides of (2.2.5) and (2.2.8), let us note a couple

of things. First of all, strictly speaking our Markov chain is not time-homogeneous.

This is not a big deal: let υm be the uniform measure on the m-cycles of Sn and υ2 be

uniform on the transpositions, then the law µk+1 of Xk+1 is given by µk+1 = υ2∗k ∗υm ,

with Fourier transform µd b2 k υc

k+1 = υ m.

Secondly, the limiting distribution of our Markov chain is not uniform on the

whole group Sn , but rather alternates between the uniform measure on An and

the uniform measure on Sn \An . This is slightly more problematic. Diaconis and

Shahshahani, as well as most of those who followed, avoided parity by making their

chain lazy. The trade-off is a small amount of precision in the ensuing computations.

We consider instead the restrictions of the representations of Sn to An . The result

is the following proposition, which we will prove at the end of the chapter:

Lemma 2.2.10. Let µ be a measure on Sn with support in An , and let U be uniform

on An . Then

X 1 X
(µ(g) − U (g))2 = dρ tr[µ̂(ρ)(µ̂(ρ))† ]. (2.2.9)
g∈Sn
n!
ρ∈Sc n
ρ6=ρtriv ,ρsign

The same holds if the support of µ is in Sn \An and U is uniform on Sn \An .

Corollary 2.2.11. With µk and Uk as defined in Chapter 1, we have the L2 equality

X 1 X
(µk+1 (g) − Uk+1 (g))2 = dρ tr[((υb2 (ρ))k υc 2
m (ρ)) ] (2.2.10)
g∈Sn
n!
ρ∈Sc n
ρ6=ρtriv ,ρsign

17
and the total variation bound

1 X
4kµk+1 − Uk+1 k2TV ≤ dρ tr[((υb2 (ρ))k υc 2
m (ρ)) ]. (2.2.11)
2
ρ∈Sc n
ρ6=ρtriv ,ρsign

Proof. Equation (2.2.10) is clear from Propositions 2.2.7-2.2.9 and Lemma 2.2.10.

To see (2.2.11), observe that µk+1 (g) − Uk+1 (g) = 0 for half of Sn , so
!2
X n! X
|µk (g) − Uk (g)| ≤ (µk+1 (g) − Uk+1 (g))2 (2.2.12)
g∈Sn
2 g∈S
n

by Cauchy-Schwarz.

If G is abelian, then any irrep of G is 1-dimensional, so that the matrices υ(ρ)

in (2.2.5) and (2.2.8) are all just scalars. Fortunately, even for a non-abelian G, a

certain type of measures on G mimics measures on abelian groups.

Definition 2.2.12. A measure υ on G is a class measure if it is constant on the

conjugacy classes of G. Note that class measures are clearly symmetric.

Lemma 2.2.13. Let υ be a class measure. For every ρ ∈ Ĝ, we have that
!
1 X
υ̂(ρ) = υ(g)χρ (g) Idρ , (2.2.13)
dρ g

where χρ (g) = tr(ρ(g)) is the character of ρ at g.

Proof. See Lemma 16.24 of [Beh00].

Remark. Since traces are similarity-invariant, χρ (g) = χρ (h) whenever g and h are

in the same conjugacy class. For elements of the symmetric group, this happens

when g and h have the same cycle type.

18
Example (Diaconis and Shahshahani, [DS81]). Consider the (lazy) random transpo-

sition shuffle on n cards, the time-homogeneous Markov chain on Sn with increment

1 2 n(n−1)
measure υ that assigns mass n
to the identity and n2
to each of the 2
transpo-

sitions τ . By Lemma 2.2.13,

 k
k 1 (n − 1)χρ (τ )
(υ̂(ρ)) = + Idρ , (2.2.14)
n ndρ

which turns (2.2.8) into

 2k
X 1 (n − 1)χρ (τ )
4kµk − U k2TV ≤ d2ρ + . (2.2.15)
n ndρ
ρ∈Scn
ρ6=ρtriv

The spectral interpretation of the right hand side of (2.2.15) is that the eigenvalues

of the transition matrix associated with the shuffle are

1 (n − 1)χρ (τ )
+ , ρ∈S
cn , (2.2.16)
n ndρ

each occuring with multiplicity d2ρ . For more on the spectral theory of Markov

chains, see Chapters 12 and 13 of [LPW08].

Analogously, for our Markov chain,

 k  
k χρ (τ ) χρ (π)
(υb2 (ρ)) υc
m (ρ) = Idρ , (2.2.17)
dρ dρ

where π is any m-cycle and τ is any transposition. Corollary 2.2.11 then gives
 2k  2
X 21 X χρ (τ ) χρ (π)
(µk+1 (g) − Uk+1 (g)) = d2ρ (2.2.18)
g∈Sn
n! dρ dρ
ρ∈Sc n
ρ6=ρtriv ,ρsign

19
and
 2k  2
1 X χρ (τ ) χρ (π)
4kµk+1 − Uk+1 k2TV ≤ d2ρ . (2.2.19)
2 dρ dρ
ρ∈Sc n
ρ6=ρtriv ,ρsign

χρ
Expressions of the form dρ
are called normalized characters. The next step is to

compute the relevant ones of these for Sn .

2.3 Representation theory of Sn

We now turn our attention to the representations and characters of Sn . For a

thorough introduction, see [Sag01] or Part I of [FH91].

Recall that Ĝ is, roughly speaking, a collection of the non-redundant irreducible

representations of G. Such a collection is in general not unique, so it would be

helpful to establish a canonical Ĝ. It is well-known (e.g. see Proposition 1.10.1 of

[Sag01]) that the number of equivalence classes of irreps is equal to the number of

conjugacy classes of G. While an explicit correspondence has not been achieved

for arbitrary groups, for Sn we can index both the conjugacy classes and the irreps

with the partitions of n. As we describe below, the partitions of n give rise to a

canonical S
cn .

Definition 2.3.1. A Young diagram of size n is a configuration of n boxes, arranged

in left-justified rows, such that the row lengths are weakly decreasing. For each

partition λ = (λ1 , λ2 , . . . , λl ) of n, the Young diagram (of shape) λ contains λi

20
boxes in its ith row.

Example. Figure 2.1 displays the Young diagrams corresponding to the partitions

of 4.

(4) (3, 1) (2, 2) (2, 1, 1) (14 )

Figure 2.1: Young diagrams corresponding to the partitions of 4

Definition 2.3.2. Let λ ` n. A Young tableau of shape λ is obtained from the

Young diagram of shape λ by filling its boxes with the numbers 1, 2, . . . , n bijec-

tively. A Young tableau is standard if the entries in each row and each column are

increasing.

At this point we shall briefly describe the construction of Specht modules, which

are indexed by partitions of n and form a complete set of irreps of Sn . See Chapter

2 of [Sag01] for the details.

Definition 2.3.3. Two Young tableaux t1 and t2 of the same shape are row equiv-

alent if corresponding rows of the two tableaux contain the same elements. For a

Young tableau t, the λ-tabloid {t} is the set of all Young tableaux that are row

equivalent to t.

A permutation σ acts on a Young tableau by replacing each number x in the

tableau with σ(x). This action gives rise to an Sn -module.

21
Definition 2.3.4. Let λ ` n. The vector space over C whose basis is the list of

λ-tabloids, denoted as M λ , is the permutation module corresponding to λ.

Definition 2.3.5. Suppose that the Young tableau t has columns C1 , C2 , . . . , Ck .

Then the column-stabilizer of t is

Ct = SC1 × SC2 × . . . × SCk , (2.3.1)

i.e. the subgroup of Sn that permutes only the elements within each column of t.

P
Definition 2.3.6. For a Young tableau t, define κt = σ∈Ct sign(σ)σ. Then the

associated polytabloid of t is given by et = κt {t}.

Definition 2.3.7. For each partition λ, the corresponding Specht module, S λ , is

the submodule of M λ spanned by all polytabloids et with t of shape λ.

Theorem 2.3.8. The Specht modules S λ for λ ` n form a complete set of irreps of

Sn over C.

Proof. See Theorem 2.4.6 of [Sag01].

n
We note here that S (n) is the trivial representation of Sn and that S (1 ) is the sign

representation of Sn . These are the only canonical 1-dimensional representations

of Sn . In general, the dimension of S λ is the number of distinct standard Young

tableaux of shape λ, which can be computed with the elegant hook length formula

of Frame, Robinson, and Thrall:

22
Definition 2.3.9. Let (i, j) denote the j th box in the ith row of a Young diagram.

Its hook is the set of all boxes directly below and directly to the right (including

itself), i.e.

Hi,j = {(i0 , j) : i0 ≥ i} ∪ {(i, j 0 ) : j 0 ≥ j}, (2.3.2)

with corresponding hook length hi,j = |Hi,j |.

Theorem 2.3.10 (Hook length formula, [FRT54]). For any partition λ of n,

n!
dim S λ = Q . (2.3.3)
(i,j)∈λ hi,j

Proof. See Theorem 3.10.2 of [Sag01].

Example. Consider the Young diagram of shape (4, 4, 3). On the left of Figure 2.2,

the dotted boxes constitute the hook H1,2 . On the right, the number in each box is

the length of the hook of the box, from which we see that the dimension of S (4,4,3)

11!
is 6·52 ·42 ·32 ·22 ·12
.

• • • 6 5 4 2
• 5 4 3 1
• 3 2 1

Figure 2.2: H1,2 and the array of hook lengths for (4, 4, 3)

In addition to being useful for finding the dimensions of representations, Young

diagrams are helpful for computing characters.

Definition 2.3.11. A rim hook ξ of a Young diagram λ is an edge-connected set of

23
boxes, containing no subset of 2-by-2 blocks, that can be removed from λ to leave

a proper Young diagram with the same top left corner as λ. The leg length of ξ,

ll(ξ), is the number of rows of ξ minus one.

Example. The top half of Figure 2.3 shows several rim hooks of (4, 4, 3) along with

their leg lengths, while the bottom half gives several non-examples of rim hooks.

• •
• • • •
• •

ll(ξ) = 0 ll(ξ) = 1 ll(ξ) = 2

• • • • • •
• • • • •
• •

Figure 2.3: Examples and non-examples of rim hooks

We use λ\ξ to denote the Young diagram obtained from λ by removing the

rim hook ξ. In the top right diagram of Figure 2.3, for instance, we have that

(4, 4, 3)\ξ = (3, 2, 1). Also, for cycle type γ = (γ1 , γ2 , . . . , γr ), we use the notation

that γ\γ1 = (γ2 , . . . , γr ). Moreover, we denote by χλγ the character of S λ on the

conjugacy class (of cycle type) γ.

Theorem 2.3.12 (Murnaghan-Nakayama rule, [Mur37] and [Nak40]). If λ is a

partition of n and γ is the cycle type of an element of Sn , then

X λ\ξ
χλγ = (−1)ll(ξ) χγ\γ1 , (2.3.4)
ξ

24
where the sum is over all rim hooks ξ of λ with γ1 boxes.

Proof. See Theorem 4.10.2 of [Sag01].

Remark. This is a recursive formula. The first iteration is to remove from λ a rim

hook with γ1 boxes in all possible ways, the next iteration is to remove from each

remaining diagram a rim hook with γ2 boxes in all possible ways, and so on. The

process terminates either when it is impossible to remove a rim hook of designated

size, so that the contribution of the corresponding character is zero, or when all

boxes have been deleted, leaving a contribution of ±1.

(4,4,3)
Example. Figure 2.4 illustrates how to compute the character χ(5,4,2) using the

Murnaghan-Nakayama rule. The sign of the rim hook being removed (±1 depending

on (−1)ll(ξ) , or 0 if no rim hook can be removed) is indicated below each diagram.

We multiply together the signs along each path and add the products, so that

(4,4,3) (4,2) (3,2,1) (1,1)

χ(5,4,2) = −χ(4,2) + χ(4,2) = (−1)2 χ(2) + 0 = (−1)3 + 0 = −1. (2.3.5)

→ • • • → •
• •
• •
• • •
−1 −1 −1
•
• • →
• •
+1 0

(4,4,3)
Figure 2.4: Computing χ(5,4,2) with the Murnaghan-Nakayama rule

25
 The stage is now set. We have the tools we need to compute the dimensions and

characters of the representations of Sn . Before we do, however, we prove Lemma

2.2.10 as promised.

Definition 2.3.13. For a partition λ, its conjugate partition, λ0 , is the partition

corresponding to the Young diagram obtained by switching the rows and columns

of λ. If λ = λ0 , then λ is said to be self-conjugate.

Example. The partitions (4, 4, 3) and (3, 3, 3, 2) are conjugates, and the partition

(4, 3, 3, 1) is self-conjugate.

0
Remark. Note that by the hook length formula, dim S λ = dim S λ . Furthermore,
0
6.6 of [Jam78] implies that χλγ = ±χλγ , depending on the sign of γ.

There is a natural correspondence between the self-conjugate partitions of n and

the conjugacy classes of Sn that split in An , which have cycle types with all odd

cycle lengths: the cycle type γ = (γ1 , γ2 , . . . , γr ) corresponds to the self-conjugate

Young diagram whose diagonal boxes have hook lengths γ1 , γ2 , . . . , γr . For instance,

the cycle type (7, 3, 1) corresponds to the partition (4, 3, 3, 1).

0
Proposition 2.3.14. (1) If λ is not self-conjugate, then S λ |An = S λ |An , and this

is irreducible as a representation of An .

(2) If λ is self-conjugate, then S λ |An = ρ1 ⊕ ρ2 , where ρ1 and ρ2 are irreps of An

dim S λ
of dimension 2
. For conjugacy classes γ of Sn that do not correspond to λ as

26
described above (even if γ also splits in An ),
λ
χSγ
χρ1 (γ) = χρ2 (γ) = . (2.3.6)
2

For the conjugacy class γ = (γ1 , γ2 , . . . , γr ) that corresponds to λ, let ξ and ξ 0 be

the classes of An that it splits into, then χρ1 (ξ) = χρ2 (ξ 0 ) and χρ1 (ξ 0 ) = χρ2 (ξ), and

furthermore these characters are given by

1 p  n−r
(−1)q ± (−1)q γ1 γ2 . . . γr , where q = . (2.3.7)
2 2

Proof. See Propositions 5.1 and 5.3 of [FH91].

n n
Proof of Lemma 2.2.10. First, observe that µ̂(S (1 ) ) and Û (S (1 ) ) are equal to 1 if

µ and U are supported on An and −1 if µ and U are supported on Sn \An , so by

(2.2.3) it suffices to show that Û (S λ ) = 0 for all λ 6= (n), (1n ).

Suppose that λ is not self-conjugate. By the first part of Proposition 2.3.14,

S λ |An is a non-trivial irrep of An , so by Proposition 2.2.9 with ρ = S λ |An and

S λ (g) = 0. But Proposition 2.2.9 also implies that

P
G = An , we have that g∈An

S λ (g) = 0, so that S λ (g) = 0 as well! Thus Û (S λ ) = 0 whether U

P P
g∈Sn g∈Sn \An

is uniform on An or on Sn \An .

Now suppose that λ is self-conjugate. The second part of Proposition 2.3.14

tells us that S λ |An = ρ1 ⊕ ρ2 , where ρ1 and ρ2 are non-trivial irreps of An . Since

S λ (g) = 0 and,
P P P
g∈An ρ1 (g) = 0 and g∈An ρ2 (g) = 0, we again have that g∈An

S λ (g) = 0.
P
analogously to above, that g∈Sn \An

27
Chapter 3

Upper Bound

Einmal ist keinmal, says Tomas to himself. What happens but once,
says the German adage, might as well not have happened at all.

– The Unbearable Lightness Of Being, “Lightness and Weight”

3.1 The m = n case

The goal of this section is to prove the following:

Theorem 3.1.1. For any c > 0, after one n-cycle and cn transpositions,

e−4c
4kµcn+1 − Ucn+1 k2TV ≤ + o(1) (3.1.1)
1 − e−4c

as n goes to infinity.

The first and most critical step of the proof is the observation that, discounting

(n) and (1n ), χλ(n) = 0 for all λ except the L-shaped ones, for which λ2 = 1. This is

an almost trivial consequence of the Murnaghan-Nakayama rule, as it is impossible

28
to remove a rim hook of size n from a Young diagram of size n unless the Young

diagram itself is the rim hook; we will discuss later what this means probabilistically.

Moreover, for an L-shaped λ, it is clear that χλ(n) is equal to 1 if λ has an odd number

of rows and −1 if λ has an even number of rows. Thus we arrive at a significant

simplication of (2.2.19), namely that

!2k
1 X χλ(2,1n−2 )
4kµk+1 − Uk+1 k2TV ≤ , (3.1.2)
2 λ∈Λ dim S λ
n

where

Λn = {λ ` n : λ1 > 1 and λ2 = 1}. (3.1.3)

χλ
(2,1n−2 )
The normalized characters dim S λ
have a simple description when λ ∈ Λn :

Proposition 3.1.2. Let λ ∈ Λn , and let j be one less than the number of rows of
 n−1 
λ. For 1 ≤ j ≤ 2
,
(n−j,1j )
χ(2,1n−2 ) n − 1 − 2j
= . (3.1.4)
dim S (n−j,1j ) n−1

Proof. By the hook length formula,

 
(n−j,1j ) n! n−1
dim S = = . (3.1.5)
n · j!(n − j − 1)! j

If j > 1, the first iteration of the Murnaghan-Nakayama rule, where we remove

a rim hook with two boxes, results in

(n−j,1j ) (n−j−2,1j ) (n−j,1j−2 )

χ(2,1n−2 ) = χ(1n−2 ) − χ(1n−2 ) . (3.1.6)

29
Let ñ be the number of remaining boxes, i.e. n − 2. Observe that, for any partition

λ̃ of ñ, the character of λ̃ at (1ñ ) is exactly the number of standard Young tableaux

of shape λ̃, or the dimension of λ̃, which again can be computed with the hook

length formula:
 
(n−j−2,1j ) (n − 2)! n−3
χ(1n−2 ) = = (3.1.7)
(n − 2) · j!(n − j − 3)! j

and
 
(n−j,1j−2 ) (n − 2)! n−3
χ(1n−2 ) = = . (3.1.8)
(n − 2) · (j − 2)!(n − j − 1)! j−2

Putting (3.1.5)-(3.1.8) together and simplifying, we get that

(n−j,1j )
χ(2,1n−2 )
 
(n − 3)! (n − 3)! j!(n − j − 1)!
(n−j,1j )
= − ·
dim S j!(n − j − 3)! (j − 2)!(n − j − 1)! (n − 1)!
(n − 3)![(n − j − 1)(n − j − 2) − j(j − 1)]
=
j!(n − j − 1)!
j!(n − j − 1)! (3.1.9)
·
(n − 1)!
n2 − 3n − 2nj + 4j + 2
=
(n − 1)(n − 2)
(n − 1 − 2j)(n − 2) n − 1 − 2j
= =
(n − 1)(n − 2) n−1

for j > 1.

For j = 1, dim S (n,1) = n − 1, and since there is only one way to remove a rim
(n−1,1)
hook of size two from (n − 1, 1), we see that χ(2,1n−2 ) = n − 3.

Remark. When λ̃ is an L-shaped partition of ñ, we can actually skip the hook length

formula and derive χλ̃(1ñ ) with the following simple combinatorial argument:

30
 Let j̃ be one less than the number of boxes in the first column of λ̃. Removing

one box at a time according to the Murnaghan-Nakayama rule, j̃ boxes in the first

column are removed before we are left with a single row of boxes, at which point

there is only one way to remove the remaining boxes. The number of ways to get

to that point is the number of ways to pace the removal of the j̃ boxes throughout

the removal of an overall ñ − 1 boxes (the upper left box must be removed last),
ñ−1


that is, j̃
.

Proof of Theorem 3.1.1. Fix any c > 0. By calculus, for n − 1 − 2j > 0,

 2cn
n − 1 − 2j
lim = e−4cj . (3.1.10)
n→∞ n−1

0
Thus Proposition 3.1.2 and the fact that χλγ = ±χλγ imply that, for large n,

 (n−2)/2
P −4cj
λ
!2cn   2 e n is even
X χ(2,1n−2 ) 

j=1
∼ (3.1.11)
λ∈Λn
dim S λ 
 (n−3)/2
P −4cj
2 e n is odd.


j=1

Summing the geometric series gives

!2cn
1 X χλ(2,1n−2 ) e−4c
4kµcn+1 − Ucn+1 k2TV ≤ ∼ , (3.1.12)
2 λ∈Λ dim S λ 1 − e−4c
n

as was to be shown.

3.2 The m = n − 1 case

Next we prove an upper bound for the m = n − 1 case.

31
Theorem 3.2.1. For any c > 0, after one (n − 1)-cycle and cn transpositions,

e−8c
4kµcn+1 − Ucn+1 k2TV ≤ + o(1) (3.2.1)
1 − e−4c

as n goes to infinity.

The proof is similar to the m = n case. We start with the observation that

χλ(n−1,1) = 0 for all λ except the ones with a 2-by-2 block of boxes in the upper left,

for which λ2 = 2 and λ3 = 0 or 1 (see Figure 3.1).

Figure 3.1: Examples of λ for which χλ(n−1,1) 6= 0

For such λ, we again have that χλ(n−1,1) = ±1, which gives

!2k
1 X χλ(2,1n−2 )
4kµk+1 − Uk+1 k2TV ≤ , (3.2.2)
2 λ∈Λ dim S λ
n−1

where

Λn−1 = {λ ` n : λ2 = 2 and λ3 = 0 or 1}. (3.2.3)

Proposition 3.2.2. Let λ ∈ Λn−1 , and let j be two less than the number of rows
 n−4 
of λ. For 0 ≤ j ≤ 2
,
(n−2−j,2,1j )
χ(2,1n−2 ) n − 4 − 2j
= . (3.2.4)
dim S (n−2−j,2,1j ) n

32
Proof. By the hook length formula,

j n!
dim S (n−2−j,2,1 ) = . (3.2.5)
(n − 1)(2 + j)(n − 2 − j) · j!(n − 4 − j)!

For j = 0, e.g. the leftmost diagram in Figure 3.1, there are two ways to remove

a rim hook of size two: from the first row, or from the second. The latter leaves a
(n−2,2)
single row and therefore contributes +1 to the value of χ(2,1n−2 ) , whereas the former

contributes

(n−4,2) (n − 2)! (n − 2)(n − 5)

χ(1n−2 ) = = . (3.2.6)
2(n − 3)(n − 4) · (n − 6)! 2

Thus
(n−2,2)
χ(2,1n−2 ) ((n − 2)(n − 5) + 2) 2(n − 1)(n − 2) · (n − 4)!
= ·
dim S (n−2,2) 2 n! (3.2.7)
n2 − 7n + 12 (n − 4)(n − 3) n−4
= = = .
n(n − 3) n(n − 3) n

For j = 1, e.g. the middle diagram in Figure 3.1, there is only one way to remove

a rim hook of size two, namely from the first row, so that

(n−3,2,1) (n−5,2,1) (n − 2)!

χ(2,1n−2 ) = χ(1n−2 ) =
3(n − 3)(n − 5) · (n − 7)!
(3.2.8)
(n − 2)(n − 4)(n − 6)
= ,
3
and
(n−3,2,1)
χ(2,1n−2 ) (n − 2)(n − 4)(n − 6) 3(n − 1)(n − 3) · (n − 5)!
= ·
dim S (n−3,2,1) 3 n! (3.2.9)
(n − 2)(n − 4)(n − 6) n−6
= = .
n(n − 2)(n − 4) n

For j > 1, there are two ways to remove a rim hook of size two: from the first

33
row, or from the first column. This implies that

(n−2−j,2,1j ) (n−4−j,2,1j ) (n−2−j,2,1j−2 )

χ(2,1n−2 ) = χ(1n−2 ) − χ(1n−2 ) , (3.2.10)

with

(n−4−j,2,1j ) (n − 2)!
χ(1n−2 ) = (3.2.11)
(n − 3)(2 + j)(n − 4 − j) · j!(n − 6 − j)!

and

(n−2−j,2,1j−2 ) (n − 2)!
χ(1n−2 ) = . (3.2.12)
j(n − 3)(n − 2 − j) · (j − 2)!(n − 4 − j)!

Combining and simplifying,

(n−2−j,2,1j ) (n − 2)![j(n − 2 − j)(n − 5 − j) − j(2 + j)(j − 1)]

χ(2,1n−2 ) =
j(n − 3)(2 + j)(n − 2 − j) · j!(n − 4 − j)!
(n − 2)!j(n − 3)(n − 4 − 2j)
= (3.2.13)
j(n − 3)(2 + j)(n − 2 − j) · j!(n − 4 − j)!
(n − 2)!(n − 4 − 2j)
= ,
(2 + j)(n − 2 − j) · j!(n − 4 − j)!

and
(n−2−j,2,1j )
χ(2,1n−2 ) (n − 2)!(n − 4 − 2j)
=
dim S (n−2−j,2,1j ) (2 + j)(n − 2 − j) · j!(n − 4 − j)!
(n − 1)(2 + j)(n − 2 − j) · j!(n − 4 − j)! (3.2.14)
·
n!
n − 4 − 2j
= ,
n
as promised.

Proof of Theorem 3.2.1. Fix any c > 0. For n − 4 − 2j > 0,

 2cn
n − 4 − 2j
lim = e−2c(4+2j) , (3.2.15)
n→∞ n

34
and thus for large n,


 (n−6)/2
P −2c(4+2j)
2 e n is even
!2cn 
χλ(2,1n−2 )
X 
j=0
∼ (3.2.16)
λ∈Λn−1
dim S λ 
 (n−5)/2
P −2c(4+2j)
2 e n is odd.


j=0

This gives
!2cn
1 X χλ(2,1n−2 ) e−8c
4kµcn+1 − Ucn+1 k2TV ≤ ∼ (3.2.17)
2 λ∈Λ dim S λ 1 − e−4c
n−1

as an upper bound.

We pause here for a few remarks. First, it is worth pointing out just how good

Theorems 3.1.1 and 3.2.1 are, in the sense that the only source of inequality comes

from Cauchy-Schwarz. This is the payoff of Lemma 2.2.10.

Secondly, the proofs of Propositions 3.1.2 and 3.2.2, while messy, are satisfying

in that only the hook length formula and the Murnaghan-Nakayama rule are used.

On the other hand, the results turned out to be essentially special cases of the

identity

χλ(2,1n−2 ) 2
P
i (λi − (2i − 1)λi )
= , (3.2.18)
dim S λ n(n − 1)

known as early as to Frobenius in [Fro00].

Thirdly, representation theory confirms what seems intuitive, that moving a lot

of cards in the beginning leads to the cards being mixed sooner. In particular, the

initial m-cycle promotes mixing by nullifying the contributions of some representa-

tions and lessening the contributions of the rest. However, we have also uncovered

35
something counterintuitive, that moving n − 1 cards in the beginning seems to lead

to even faster mixing than moving all n cards! We will verify this and propose an

explanation as we tackle the lower bound.

36
Chapter 4

Fixed Points and Lower Bound

A question is like a knife that slices through the stage backdrop and
gives us a look at what lies hidden behind it.

– The Unbearable Lightness Of Being, “The Grand March”

4.1 Fixed points

For measures µ and ν on a set G, one of the classic approaches to finding a lower

bound for kµ − νkTV is to identify a subset A of G where |µ(A) − ν(A)| is close

to maximal. In many mixing problems involving the symmetric group, it is conve-

nient to make A either the set of fixed-point-free permutations or its complement,

since it is well-known (e.g. to Montmort three centuries ago in [Mon08]) that the

distribution of the number of fixed points with respect to the uniform measure on

Sn is asymptotically P(1), the Poisson distribution of mean one.

37
 Slightly less well-known6 , though unsurprising, is that the distribution of fixed

points with respect to the uniform measure on either An or Sn \An is also asymptot-

ically P(1). We will give an original proof for all of the Poisson limit laws mentioned

here in Section 4.3. For a brute-force combinatorial proof of the weaker result that

the mass, with respect to the uniform measure on Sn , An , as well as Sn \An , of fixed-

1
point-free permutations approaches e
as n approaches infinity, consult [AU08].

For Diaconis and Shahshahani’s random transposition shuffle, A is the set of

permutations with one or more fixed points, and finding µk (A) boils down to a

coupon collector’s problem. Let B be the event that, after k transpositions, at least

one card is untouched. It is not difficult to see that µk (A) ≥ P(B), where P(B)

is equal to the probability that at least one of n coupons is still missing after 2k

trials. The coupon collector’s problem is well-studied (see, for instance, Section

IV.2 of [Fel68]), so this immediately gives a lower bound for µk (A), which in turn7

produces a lower bound for kµk (A) − U (A)kTV .

The above argument is so short and simple that it was tagged on to the end of

the introduction of [DS81], as if an afterthought. Unfortunately, it is completely

inapplicable to our problem, since the initial (large) m-cycle obliterates the core of

the argument. To delve more deeply into the behavior of fixed points, we again

turn to representation theory.

6
We have not actually found this documented anywhere but presume that it is known.
7
As µk (A) − U (A) ≥ 0, the inequality is in the desired direction.

38
4.2 The defining representation

Definition 4.2.1. The defining, or permutation, representation of Sn is the n-

dimensional representation % where (%(σ))i,j is 1 if σ(j) = i and 0 otherwise.

Example. For S3 ,
   
1 0 0 0 1 0
%(e) =  0 1 0  , %(1, 2) =  1 0 0  ,
0 0 1 0 0 1

   
0 0 1 1 0 0
%(1, 3) =  0 1 0  , %(2, 3) =  0 0 1  , (4.2.1)
1 0 0 0 1 0

   
0 0 1 0 1 0
%(1, 2, 3) =  1 0 0  , %(1, 3, 2) =  0 0 1  .
0 1 0 1 0 0

The significance of % should be apparent: the fixed points can be read off of

the matrix diagonal, so that χ% (σ) is precisely the number of fixed points of σ. We

should also point out that % is reducible and decomposes as S (n−1,1) ⊕ S (n) (see

Examples 1.4.3, 1.9.5, and 2.3.8 of [Sag01]), so that the character of S (n−1,1) at σ is

one less than the number of fixed points of σ. The representation S (n−1,1) is often

referred to as the standard representation of Sn .

Heuristically, the connection between S (n−1,1) and fixed points vouches for the

quality of the lower bound obtained via fixed points, since S (n−1,1) is in some sense

the representation closest to the trivial representation and usually contributes the

largest normalized character to the sum in (2.2.19). Moreover, this connection sheds

39
light on why the m = n − 1 case seems to converge faster: it is an atypical case

where the contribution from S (n−1,1) is zero! Informally, this is the representation-

theoretic analogue of the probabilistic intuition that, since the expected number of

fixed points is one under the uniform distribution, a chain that starts with exactly

one fixed point is closer to uniformity than a chain that starts with none.

Now, heuristics aside, we would like to find the mass of fixed-point-free permuta-

tions under µk+1 . Since this cannot be done directly, we will in fact prove something

more general: we will fully characterize the distribution of χ% with respect to µk+1

by deriving all moments of χ% with respect to µk+1 . The pivotal observation, in-

spired by Remark 1 in Chapter 3D of [Dia88], is the following, which relates raw

moments of the fixed point distribution to tensor powers of %:

Proposition 4.2.2. Let Eµ denote expectation with respect to µ, and let aλ,r be

the multiplicity of S λ in the decomposition of %⊗r into a direct sum of irreducible

representations, i.e. let

M M
%⊗r = aλ,r S λ := (S λ )⊕aλ,r . (4.2.2)
λ`n λ`n

Then, viewing χ% as a random variable on Sn ,

X
Eµ ((χ% )r ) = aλ,r tr(µ̂(S λ )) (4.2.3)
λ`n

for any positive integer r.

Proof. Since the tensor product has the property that the trace of the product is

40
equal to the product of the traces,
X X
Eµ ((χ% )r ) = µ(σ)[tr(%(σ))]r = µ(σ)tr(%⊗r (σ))
σ∈Sn σ∈Sn
! !
X M X X
= µ(σ)tr aλ,r S λ (σ) = µ(σ) aλ,r tr(S λ (σ)) (4.2.4)
σ∈Sn λ`n σ∈Sn λ`n
!
X X X
= aλ,r µ(σ)tr(S λ (σ)) = aλ,r tr(µ̂(S λ )),
λ`n σ∈Sn λ`n

where the last equality is due to the linearity of the trace.

Remark. The first line of (4.2.4) is clearly true for any representation ρ of Sn , and

hence the equality Eµ ((χρ )r ) = tr(µ̂(ρ⊗r )) holds for all ρ.

Recall that by (2.2.17),

!k
λ
χλ(2,1n−2 )
tr(µd b2 (S λ ))k υc
k+1 (S )) = tr[(υ
λ λ
m (S )] = χ(m,1n−m ) , (4.2.5)
dim S λ

which we have already computed for all λ in the m = n and m = n − 1 cases while

working on the upper bound! Thus in light of Proposition 4.2.2, if we find aλ,r for

all λ and r, then we would know all moments of χ% with respect to µk+1 .

We shall do just that.

Theorem 4.2.3. Let λ ` n and 1 ≤ r ≤ n − λ2 . The multiplicity of S λ in the irrep

decomposition of %⊗r is given by

r   
X
λ̄ i r
aλ,r =f , (4.2.6)
|λ̄| i
i=|λ̄|

where λ̄ is the truncated partition (λ2 , . . . , λl ) of weight |λ̄|, f λ̄ is the number of

r
standard Young tableaux of shape λ̄, and i
is a Stirling number of the second

41
kind, i.e. the number of ways to partition r objects into i non-empty subsets.

Remark. Since f λ̄ = dim S (λ2 ,...,λl ) can be computed with the hook length formula

and the Stirling numbers can be explicitly defined as

  i  
r 1X i−j i
= (−1) jr, (4.2.7)
i i! j=1 j

we can rewrite (4.2.6) as

r   i   !
n! X i 1X i−j i
aλ,r = Q (−1) jr , (4.2.8)
(x,y)∈λ hx,y |λ̄| i! j=1 j
i=|λ̄|

which defines aλ,r in terms of elementary expressions and factorials.

Proof. Theorem 4.2.3 owes its existence to the recent work of Goupil and Chauve,

who derived in [GC06] the generating function

X xr f λ̄ ex −1 x
aλ,r = e (e − 1)|λ̄| (4.2.9)
r! |λ̄|!
r≥|λ̄|

for λ ` n and n ≥ r + λ2 .

By (24b) and (24f) in Chapter 1 of [Sta97],

X  s  xs (ex − 1)j
= (4.2.10)
s≥j
j s! j!

and

X xt x
Bt = ee −1 , (4.2.11)
t≥0
t!

Pt t
where B0 := 1 and Bt = q=1 q
is the t-th Bell number, so we obtain from (4.2.9)

42
that

aλ,r X Bt  s 
λ̄
=f , (4.2.12)
r! s+t=r
s!t! |λ̄|

and thus
r−|λ̄|
aλ,r X rr − t
= Bt
f λ̄ t=0
t |λ̄|
  r−| Xλ̄| X t    
r t r r−t
= + (4.2.13)
|λ̄| t=1 q=1
q t |λ̄|
  r−|
Xλ̄| r−|
Xλ̄|  t rr − t
r
= + .
|λ̄| q=1 t=q
q t |λ̄|

By (24.1.3, II.A) of [AS65],

r−|λ̄|       
X t r r−t q + |λ̄| r
= , (4.2.14)
t=q
q t |λ̄| |λ̄| q + |λ̄|

so that
  r−|
Xλ̄| q + |λ̄| r 
aλ,r r
= +
f λ̄ |λ̄| q=1
|λ̄| q + |λ̄|
  r    X r    (4.2.15)
r X i r i r
= + = ,
|λ̄| |λ̄| i |λ̄| i
i=|λ̄|+1 i=|λ̄|

as we rejoice.

On a side note, let bλ,r be the multiplicity of S λ in the irrep decomposition of

(S (n−1,1) )⊗r , so that

M M
(S (n−1,1) )⊗r = bλ,r S λ := (S λ )⊕bλ,r . (4.2.16)
λ`n λ`n

43
Goupil and Chauve also derived the generating function

X xr f λ̄ ex −x−1 x
bλ,r = e (e − 1)|λ̄| , (4.2.17)
r! |λ̄|!
r≥|λ̄|

so from Theorem 4.2.3 we can obtain a decent formula for the irrep decomposition

of (S (n−1,1) )⊗r as well.

Corollary 4.2.4. Let λ ` n and 1 ≤ r ≤ n − λ2 . The multiplicity of S λ in the

irrep decomposition of (S (n−1,1) )⊗r is given by

 
r   X s   
X r  i s 
bλ,r = f λ̄ (−1)r−s . (4.2.18)
s |λ̄| i
s=|λ̄| i=|λ̄|

Proof. Comparing (4.2.17) with (4.2.9) gives

    !
r s s X (−x)t
X x X x X x
bλ,r =  aλ,s  e−x =  aλ,s  , (4.2.19)
r! s! s! t≥0
t!
r≥|λ̄| s≥|λ̄| s≥|λ̄|

so that
 
t r r−s s   
bλ,r X (−1) aλ,s X (−1) f λ̄
X i s 
= = , (4.2.20)
r! s+t=r
s!t! s!(r − s)! |λ̄| i
s=|λ̄| i=|λ̄|

and the result follows.

Remark. Corollary 4.2.4 is very similar to Proposition 2 of [GC06], but our result is

a bit cleaner, as it does not involve associated Stirling numbers of the second kind.

44
4.3 Lower bound for the m = n case

Before unleashing the power of Theorem 4.2.3, we need to clear up a technicality:

not all probability distributions are uniquely determined by their moments. For

instance, a distribution all of whose moments match those of the log-normal is

not necessarily log-normal. Fortunately, there is a simple sufficient condition for

uniqueness.

Theorem 4.3.1. Let mr denote the r-th moment of the distribution of a random
r
variable Y . If the moment-generating function E(etY ) = mr tr! has a positive
P
r≥0

radius of convergence, then there is no other distribution with the same moments.

Proof. See Theorem 30.1 of [Bil95].

In Theorems 4.3.3 and 4.3.4, we will argue that a sequence of distributions

converges to a Poisson. By definition, the moment-generating function for a Poisson

of mean ν is

X ν j e−ν X (et ν)j t t

etj = e−ν = e−ν ee ν = eν(e −1) , (4.3.1)
j≥0
j! j≥0
j!

which satisfies the uniqueness condition, so Poisson distributions are indeed deter-

mined by their moments.

Theorem 4.3.2. Suppose that the distribution of Y is determined by its moments,

that Y has moments of all orders, and that E(Yir ) tends to E(Y r ) for all r, then Yi

converges in distribution to Y .

45
Proof. See Theorem 30.2 of [Bil95].

We are now ready to prove several Poisson limit laws. First, as promised, we

give a new proof for an ancient result:

Theorem 4.3.3. (1) Let USn denote the uniform measure on Sn . As n approaches

infinity, the distribution of the number of fixed points of a permutation randomly

chosen according to USn converges to P(1).

(2) The above statement holds if we replace Sn with either An or Sn \An .

λ
Proof. For (1), recall Proposition 2.2.9, which implies that U
dSn (S ) is 1 if λ = (n)

and 0 otherwise. Thus the combination of Proposition 4.2.2 and Theorem 4.2.3

gives that, for 1 ≤ r ≤ n,

r  
r
X r
EUSn ((χ% ) ) = a(n),r = = Br , (4.3.2)
i=0
i

which by (4.2.11) and (4.3.1) is exactly the r-th moment of P(1). This means that

the first n moments of χ% with respect to USn match those of P(1), and therefore

convergence follows from Theorem 4.3.2.

λ n
For (2), recall from the proof of 2.2.10 that U
d An (S ) is 1 if λ is (n) or (1 ) and

λ n
Sn \An (S ) is 1 if λ = (n), −1 if λ = (1 ), and 0 otherwise.
0 otherwise. Moreover, U\

Hence

EUAn ((χ% )r ) = a(n),r + a(1n ),r

(4.3.3)
and EUSn \An ((χ% )r ) = a(n),r − a(1n ),r .

46
As before, a(n),r = Br for 1 ≤ r ≤ n. Meanwhile, for 1 ≤ r ≤ n − 1,

r   
X i r
a(1n ),r = , (4.3.4)
i=n−1
n − 1 i

which is 0 for 1 ≤ r ≤ n − 2. Thus the first n − 2 moments of χ% with respect to

either UAn or USn \An match those of P(1).

Returning to the Markov chain mixing rate problem, the next Poisson limit law

will finally give a satisfactory lower bound for the mixing rate of the n-cycle-to-

transpositions shuffle.

Theorem 4.3.4. Fix any c > 0. As n approaches infinity, the distribution of

the number of fixed points after one n-cycle and cn transpositions converges to

P(1 − e−2c ).

Proof. One can deduce from the moment-generating function, or just look up in
Pr r
[Rio37], that the r-th moment of P(ν) is i=1 i
ν i . As it went with the proof of

(n)
Theorem 4.3.3, µ[
cn+1 (S ) = 1, and we will ignore the alternating representation

because it suffices to consider the first n − 2 moments. For the non-trivial and

non-alternating representations, we take advantage of previous computations and

synthesize Proposition 3.1.2, (3.1.10) with n instead of 2n, and (4.2.5) with the

recollection that χλ(n) = (−1)|λ̄| to obtain



(−1)|λ̄| e−2c|λ̄| λ ∈ Λn


λ
cn+1 (S ) ∼
µ[ (4.3.5)


0

otherwise.

47
 By Theorem 4.2.3 (second line below) and (4.3.5) (fourth line), for 1 ≤ r ≤ n−2,
X
Eµcn+1 ((χ% )r ) = a(n),r + aλ,r µ[ λ
cn+1 (S )
λ∈Λn
r   n−2 X
r   
X r X r i λ
= + µ[
cn+1 (S )
i=1
i i |λ̄|
|λ̄|=1 i=|λ̄|
r   r Xi   
X r X r i λ
= + µ[
cn+1 (S )
i=1
i i=1
i |λ̄|
|λ̄|=1
r   r X i    (4.3.6)
X r X r i
∼ + (−e−2c )|λ̄|
i=1
i i=1 |λ̄|=1
i |λ̄|
 
r   i  
X r  X i
= 1+ (−e−2c )|λ̄| 
i=1
i |λ̄|
|λ̄|=1
r
X r  
= (1 − e−2c )i .
i=1
i

This shows that the first n − 2 moments of χ% with respect to µcn+1 approach those

of P(1 − e−2c ), and once again convergence follows from Theorem 4.3.2.

Corollary 4.3.5. For any c > 0, after one n-cycle and cn transpositions,

e−2c
kµcn+1 − Ucn+1 kTV ≥ − o(1) (4.3.7)
e

as n goes to infinity.

Proof. Let A be the set of fixed-point-free permutations. Then

kµcn+1 − Ucn+1 kTV ≥ |µcn+1 (A) − Ucn+1 (A)|

(4.3.8)
1 −2c (e−2c )2 e−2c
 
e−2c −1 1
∼e − = e + + ··· ≥ ,
e e 2! e

as was to be shown.

48
Remark. Together with Theorem 3.1.1, we have that

e−2c e−2c
− o(1) ≤ kµcn+1 − Ucn+1 kTV ≤ √ + o(1) (4.3.9)
e 2 1 − e−4c

as n goes to infinity. The gap is especially respectable if e−4c is small. Also, recall

from Theorem 3.2.1 that an upper bound for the m = n − 1 case is

e−4c
kµcn+1 − Ucn+1 kTV ≤ √ + o(1), (4.3.10)
2 1 − e−4c

which is smaller than even the lower bound for the m = n case so long as c is

at least approximately 0.262. Hence we can reasonably say that starting with an

(n − 1)-cycle is indeed more beneficial for mixing than starting with an n-cycle.

49
Chapter 5

Further Considerations

[V ]ertigo is something other than the fear of falling. It is the voice

of the emptiness below us which tempts and lures us, it is the desire
to fall, against which, terrified, we defend ourselves.

– The Unbearable Lightness of Being, “Soul and Body”

5.1 Miscellaneous results

In this section we use S (n−1,1) to derive two more results about expected numbers of

fixed points. Recall that the character of S (n−1,1) at σ is one less than the number

of fixed points of σ.

First, we present the following martingale-like property for Markov chains on Sn

whose increment distributions are class measures: if a chain starts with one fixed

point, then it will always average exactly one fixed point.

Proposition 5.1.1. Let X0 be the identity, and set X1 = τ1 X0 , where τ1 is selected

according to any class measure supported on the set of permutations with one fixed

50
point. For k ≥ 2, set Xk = τk Xk−1 , where τk is selected according to any class

measure on Sn (the measure can be different for each k). Then the expected number

of fixed points of Xk is one for all k ≥ 1.

Proof. Let ν1 be a class measure supported on the set of permutations with one

fixed point, ν2 , ν3 , . . . , νk be class measures on Sn , and define µk = νk ∗ · · · ∗ ν2 ∗ ν1 .

By the remark following Proposition 4.2.2,

µk (S (n−1,1) )]
Eµk (χS (n−1,1) ) = tr[c
(5.1.1)
= tr[νb1 (S (n−1,1) )νb2 (S (n−1,1) ) · · · νbk (S (n−1,1) )],

where
!
1 X
νb1 (S (n−1,1) ) = ν1 (σ)χS (n−1,1) (σ) In−1 (5.1.2)
n − 1 σ∈S
n

by Lemma 2.2.13. Consider the anatomy of the partition (n − 1, 1): under the

Murnaghan-Nakayama rule, the only way for a single box to remain at the end is

for the box in the second row to have been removed as a singleton, which requires

a cycle type with at least two fixed points. This means that χS (n−1,1) (σ) = 0 if σ

has exactly one fixed point. On the other hand, if σ does not have exactly one

fixed point, then ν1 (σ) = 0. Thus νb1 (S (n−1,1) ) = 0, which in turn implies that

Eµk (χS (n−1,1) ) = 0, and hence the expected number of fixed points with respect to

µk is one for all k ≥ 1.

Returning to the m-cycle-to-transpositions chain, we can now characterize the

expected number of fixed points for the general case where m is defined by an

51
arbitrary function m(n) of n.

Theorem 5.1.2. After one m(n)-cycle and k transpositions,



 ∼ 1 − e−2c m(n) = n, k = cn






= 1 m(n) = n − 1, any k
Eµk+1 (χ% ) (5.1.3)



∼ 1 + e−2c m(n) 6= n, n − 1,





k = cn + n2 ln(n − m(n) − 1),


where, as in Chapter 4, % = S (n−1,1) ⊕ S (n) .

Proof. The m(n) = n (Theorem 4.3.4) and m(n) = n−1 (special case of Proposition

5.1.1) cases have already been shown. For m(n) 6= n, n − 1,

(n−1,1)
χ(m(n),1n−m(n) ) = n − m(n) − 1, (5.1.4)

so by (4.2.5),
(n−1,1) !k
(n−1,1)
χ(2,1n−2 )
Eµk (χS (n−1,1) ) = χ(m(n),1n−m(n) )
dim S (n−1,1)
(5.1.5)
 k
n−3
= (n − m(n) − 1) .
n−1
n
Setting k = cn + 2
ln(n − m(n) − 1), we have that
 cn+ n2 ln(n−m(n)−1)
n−3
lim Eµk (χS (n−1,1) ) = lim (n − m(n) − 1)
n→∞ n→∞ n−1
(5.1.6)
=(n − m(n) − 1)e−2c e− ln(n−m(n)−1) = e−2c ,

and the result follows.

52
5.2 Open questions

We conclude with a list of open questions related to our work.

Question (1). What is the lower bound for the mixing time of the m = n − 1 case

of the m-cycle-to-transpositions chain?

The m = n − 1 case is trickier than the m = n case because, unlike the m = n

case, the distribution of the number of fixed points is not quite Poisson. Indeed,

after an initial (n − 1)-cycle, the expected number of fixed points is always one.

On the other hand, we can compute from either Corollary 4.3.5 or, as a more fun

exercise, Proposition 1 of [GC06] that

S (n−1,1) ⊗ S (n−1,1) = S (n) ⊕ S (n−1,1) ⊕ S (n−2,2) ⊕ S (n−2,1,1) , (5.2.1)

which along with Proposition 3.2.2 implies that the variance of the number of fixed

points after one (n−1)-cycle and cn transpositions is asymptotically 1−e−4c . As the

mean does not match the variance, the distribution is not Poisson. Nevertheless, it

must be very close to Poisson, and one may be able to compute a few more moments

and use brute force to bound the mass of fixed-point-free permutations, which in

turn will give a lower bound for the mixing time.

In general, when finding a lower bound for the mixing time of a Markov chain, the

method of moments is powerful because it produces robust results without relying

on convenient but narrowly-scoped combinatorial arguments. Theorem 4.2.3, in

particular, enables a wide class of Markov chains on Sn to be analyzed this way. On

53
the other hand, we got lucky with Theorems 4.3.3 and 4.3.4 in the sense that we

happened to recognize each sequence of moments as that of a Poisson. When the

moments do not match up with those of any well-known distributions, there is the

additional task of extracting information about the distribution from its moments.8

Question (2). In the general case where m is an arbitrary function m(n) of n

(excluding n and n − 1), is n ln(n − m(n) − 1) the correct order of mixing time?

This question is motivated by Theorem 5.1.2. To see that it is at least plausible,

consider that O(n ln(n − m(n) − 1)) is the right mixing time for m(n) = 2, i.e.

the random transposition shuffle. Proving a general upper bound with only the

techniques from this thesis is likely difficult, but the order of a lower bound may be

within reach. In particular, from (5.2.1) we should be able to completely charaterize

the variance of the number of fixed points for arbitrary m(n) like Theorem 5.1.2 did

for the expected value. If we have both the first and the second moments, then we

may be able to derive the order of a lower bound using Chebyshev’s inequality or

Proposition 7.8 of [LPW08], a method of procuring lower bounds from distinguishing

statistics.

Question (3). For a Markov chain on Sn whose increment distributions are class

measures, what conditions are sufficient for its fixed point distribution to be asymp-

totically (in n) Poisson?

8
The classical moment problem is oft-studied, but predominantly for determinancy conditions,
and most of the work on reconstruction has been for continuous distributions. See the introduction
of [MH09] for a survey of results.

54
 A necessary condition appears to be that the initial step does not create exactly

one fixed point. Is it also sufficient? By simply playing around with (5.2.1), we may

be able to identify a class of Markov chains whose fixed point distributions have

asymptotically the same mean and variance, which would be a small step toward

proving Poisson-ness but worthwhile heuristic evidence nonetheless.

Question (4). What is the contribution, if any, of Theorem 4.2.3 and Corollary

4.2.4 to related topics in algebraic combinatorics?

In particular, can Theorem 4.2.3 and Corollary 4.2.4 shed any insight on the

notoriously difficult-to-compute Kronecker coefficients? Kronecker coefficients are

the multiplicities in the tensor product decomposition of two irreps; see [BI08] for

a survey of the subject as well as a complexity-theoretic implication. Decomposi-

tions of higher tensor powers are related to plethysms of symmetric functions, and

plethystic computations have led to remarkable recent advances in the theory of

Macdonald polynomials. See Appendix 2 and Exercise 7.74 following the Chapter 7

of [Sta99] for an introduction to plethysms and [LR11] for their connections to the

Macdonald polynomials, connections that delve into some of the deepest and most

active areas of algebraic combinatorics.

We have now ventured into a field of intricately connected ideas with much

potential for further exploration. Any of these topics is sure to lead us down a

wondrous rabbit hole. However, to quote Dostoevsky, that might be the subject of

a new story, but our present story is ended.

55
Bibliography

[AS65] M. Abramowitz and I. A. Stegun, eds., Handbook of Mathematical Func-

tions, Dover, New York, 1965.

[Ald99] D. J. Aldous, Deterministic and stochastic models for coalescence (aggre-

gation and coagulation): a review of the mean-field theory for probabilists,

Bernoulli 5 (1999), no. 1, 3-48.

[AD86] D. J. Aldous and P. Diaconis, Shuffling cards and stopping times, Amer.

Math. Monthly 93 (1986), no. 5, 333-348.

[AU08] B. Ali and A. Umar, Some combinatorial properties of the alternating

group, Southeast Asian Bull. Math. 32 (2008), no. 5, 823-830.

[ADS11] S. Assaf, P. Diaconis, and K. Soundararajan, A rule of thumb for riffle

shuffling, Ann. Appl. Probab. 21 (2011), no. 3, 843-875.

[BD92] D. Bayer and P. Diaconis, Trailing the dovetail shuffle to its lair, Ann.

Appl. Probab. 2 (1992), no. 2, 294-313.

56
[Beh00] E. Behrends, Introduction to Markov Chains (with Special Emphasis on

Rapid Mixing), Vieweg Verlag, Braunschweig/Wiesbaden, 2000.

[Bil95] P. Billingsley, Probability and Measure, 3rd ed., Wiley, New York, 1995.

[BI08] P. Bürgisser and C. Ikenmeyer, The complexity of computing Kronecker

coefficients, FPSAC 2008, DMTCS proc. AJ (2008), 357-368.

[CST08] T. Ceccherini-Silberstein, F. Scarabotti, and F. Tolli, Harmonic Analy-

sis on Finite Groups: Representation Theory, Gelfand Pairs and Markov

Chains, Cambridge Stud. Adv. Math. 108, Cambridge Univ. Press, Cam-

bridge, 2008.

[CH10] M. A. Conger and J. A. Howald, A better way to deal the cards, Amer.

Math. Monthly 117 (2010), no. 8, 686-700.

[CV06] M. A. Conger and D. Viswanath, Riffle shuffles of decks with repeated

cards, Ann. Prob. 34 (2006), no. 2, 804-819.

[Dia88] P. Diaconis, Group Representations in Probability and Statistics, IMS

Lecture Notes Monogr. Ser. 11, Inst. Math. Statist., Hayward, CA, 1988.

[Dia01] P. Diaconis, Mathematical developments from the analysis of riffle shuf-

fling, in Groups, Combinatorics and Geometry: Durham 2001 (A. A.

Ivanov, M. W. Liebeck, and J. Saxl., eds.), 73-97, World Scientific, River

Edge, NJ, 2001.

57
[DF09] P. Diaconis and J. Fulman, Carries, shuffling, and symmetric functions,

Adv. in Appl. Math. 43 (2009), no.2, 176-196.

[DFH13] P. Diaconis, J. Fulman, and S. P. Holmes, Analysis of casino shelf shuf-

fling machines, Ann. Appl. Probab. 23 (2013), no. 4, 1692-1720.

[DPR12] P. Diaconis, C. Y. A. Pang, and A. Ram, Hopf algebras and Markov

chains: two examples and a theory, preprint, 2012.

[DR12] P. Diaconis and A. Ram, A probabilistic interpretation of the Macdonald

polynomials, Ann. Prob. 40 (2012), no. 5, 1861-1896.

[DS81] P. Diaconis and M. Shahshahani, Generating a random permutation with

random transpositions, Z. Wahrsch. verw. Geb. 57 (1981), no. 2, 159-179.

[Dur03] R. Durrett, Shuffling chromosomes, J. Theor. Probab. 16 (2003), no. 3,

725-750.

[Fel68] W. Feller, An Introduction to Probability Theory and Its Applications,

Vol. I, 3rd ed., Wiley, New York, 1968.

[Ful00] J. Fulman, Semisimple orbits of Lie algebras and card shuffling measures

on Coxeter groups, J. Algebra 224 (2000), no. 1, 151-165.

[Ful01] J. Fulman, Applications of the Brauer complex: card shuffling, permu-

tation statistics, and dynamical systems, J. Algebra 243 (2001), no. 1,

96-122.

58
[FH91] W. Fulton and J. Harris, Representation Theory: A First Course, GTM

129, Springer-Verlag, New York, 1991.

[Fro00] F. G. Frobenius, Über die Charaktere der symmetrischen Gruppen, Sitz.

Konig. Preuss. Akad. Wissen. (1900), 516-534.

[FRT54] J. S. Frame, G. de B. Robinson, and R. M. Thrall, The hook graphs of

the symmetric group, Canad. J. Math. 6 (1954), 316-325.

[GC06] A. Goupil and C. Chauve, Combinatorial operators for Kronecker powers

of representations of Sn , Sem. Lothar. Combin. 54 (2006), B54j.

[Hil92] M. V. Hildebrand, Generating random elements in SLn (Fq ) by random

transvections, J. Alg. Comb. 1 (1992), no. 2, 133-150.

[HMR12] V. T. Hoang, B. Morris, and P. Rogaway, An enciphering scheme

based on a card shuffle, in Advances in Cryptology – CRYPTO 2012

(R. Safavi-Naini and R. Canetti, eds.), LNCS 7417, 1-13, Springer,

Berlin/Heidelberg, 2012.

[HLW06] S. Hoory, N. Linial, and A. Wigderson, Expander graphs and their appli-

cations, Bull. Amer. Math. Soc. 43 (2006), no. 4, 439-561.

[Jam78] G. D. James, The Representation Theory of the Symmetric Groups, LNM

682, Springer-Verlag, Berlin, 1978.

59
[Jon06] J. Jonasson, The overhand shuffle mixes in Θ(n2 ln n) steps, Ann. Appl.

Probab. 16 (2006), no. 1, 231-243.

[LPW08] D. A. Levin, Y. Peres, and E. L. Wilmer, Markov Chains and Mixing

Times, American Mathematical Society, Providence, RI, 2008.

[LR11] N. A. Loehr and J. B. Remmel, A computational and combinatorial exposé

of plethystic calculus, J. Alg. Comb. 33 (2011), no. 2, 163-198.

[Lul96] N. Lulov, Random Walks on the Symmetric Group Generated by Conju-

gacy Classes, Ph.D. Thesis, Harvard University, 1996.

[Mar06] A. A. Markov, Extension of the law of large numbers to dependent events,

Bull. Soc. Math. Kazan 2 (1906), 135-156.

[Mir02] I. Mironov, (Not so) random shuffles of RC4, in Advances in Cryptol-

ogy – CRYPTO 2002 (M. Yung, ed.), LNCS 2442, 304-319, Springer,

Berlin/Heidelberg, 2002.

[MH09] R. M. Mnatsakanov and A. S. Hakobyan, Recovery of distributions via

moments, IMS Lecture Notes Monogr. Ser. 57, 252-265, Inst. Math.

Statist., Beachwood, OH, 2009.

[Mon08] P. R. de Montmort, Essay d’Analyse sur les Jeux de Hazard, Jacques

Quillau, Paris, 1708, reprinted by Chelsea, New York, 1980.

60
[MNP12] B. Morris, W. Ning, and Y. Peres, Mixing time of the card-cyclic-to-

random shuffle, preprint, 2012.

[MPS04] E. Mossel, Y. Peres, and A. Sinclair, Shuffling by semi-random transposi-

tions, Proceedings of the 45th Annual IEEE Symposium on Foundations

of Computer Science (2004), 572-581.

[Mur37] F. E. Murnaghan, The characters of the symmetric group, Amer. J. Math.

59 (1937), 739-753.

[Nak40] T. Nakayama, On some modular properties of irreducible representations

of a symmetric group, I and II, Jap. J. Math. 17 (1940), 165-184 and

411-423.

[Pem89] R. Pemantle, Randomization time for the overhand shuffle, J. Theor.

Probab. 2 (1989), no. 1, 37-49.

[Pem94] R. Pemantle, A shuffle that mixes sets of any fixed size much faster than

it mixes the whole deck, Rand. Struct. Alg. 5 (1994), no. 5, 609-625.

[Poi12] H. Poincaré, Calcul des Probabilités, 2nd ed., Gauthier-Villars, Paris,

1912.

[Rio37] J. Riordan, Moment recurrence relations for binomial, poisson and hy-

pergeometric frequency distributions, Ann. Math. Stat. 8 (1937), no. 2,

103-111.

61
[Sag01] B. E. Sagan, The Symmetric Group: Representations, Combinatorial Al-

gorithms, and Symmetric Functions, 2nd ed., GTM 203, Springer-Verlag,

New York, 2001.

[S-C01] L. Saloff-Coste, Probability on groups: random walks and invariant dif-

fusions, Notices Amer. Math. Soc. 48 (2001), no. 9, 968-977.

[S-C04] L. Saloff-Coste, Random walks on finite groups, in Probability on Dis-

crete Structures (H. Kesten, ed.), Encyclopaedia Math. Sci. 110, 263-346,

Springer-Verlag, Berlin, 2004.

[SG89] D. Sankoff and M. Goldstein, Probabilistic models of genome shuffling,

Bull. Math. Biol. 51 (1989), no. 1, 117-124.

[Sta97] R. P. Stanley, Enumerative Combinatorics, Vol. I, Wadsworth, Monterey

CA, 1986, Cambridge Stud. Adv. Math. 49, reprinted by Cambridge Univ.

Press, Cambridge, 1997.

[Sta99] R. P. Stanley, Enumerative Combinatorics, Vol. II, Cambridge Stud. Adv.

Math. 62, Cambridge Univ. Press, Cambridge, 1999.

[Sta01] R. P. Stanley, Generalized riffle shuffles and quasisymmetric functions,

Ann. Comb. 5 (2001), no. 3-4, 479-491.

[Tao12] T. Tao, Topics in Random Matrix Theory, American Mathematical So-

ciety, Providence, RI, 2012.

62
[Tho73] E. O. Thorp, Nonrandom shuffling with applications to the game of Faro,

J. Amer. Statist. Assoc. 68 (1973), no. 344, 842-847.

[Wil04] D. B. Wilson, Mixing times of Lozenge tiling and card shuffling Markov

chains, Ann. Appl. Probab. 14 (2004), no. 1, 274-325.

63