Partitions Hypergeometric Systems and

Dirichlet Processes in Statistics

SpringerBriefs in Statistics Mano
Shuhei
Visit to download the full and correct content document:
https://ebookmeta.com/product/partitions-hypergeometric-systems-and-dirichlet-proce
sses-in-statistics-springerbriefs-in-statistics-mano-shuhei-2/
More products digital (pdf, epub, mobi) instant
download maybe you interests ...

Partitions Hypergeometric Systems and Dirichlet

Processes in Statistics SpringerBriefs in Statistics
Mano Shuhei

https://ebookmeta.com/product/partitions-hypergeometric-systems-
and-dirichlet-processes-in-statistics-springerbriefs-in-
statistics-mano-shuhei-2/

Introduction to Mathematical Statistics (What's New in

Statistics) 8th Edition Robert Hogg

https://ebookmeta.com/product/introduction-to-mathematical-
statistics-whats-new-in-statistics-8th-edition-robert-hogg/

Basic Statistics in Business and Economics (ISE HED

IRWIN STATISTICS) 10th Edition Douglas A. Lind

https://ebookmeta.com/product/basic-statistics-in-business-and-
economics-ise-hed-irwin-statistics-10th-edition-douglas-a-lind/

Martingale Methods in Statistics Chapman Hall CRC

Monographs on Statistics and Applied Probability 1st
Edition Nishiyama Yoichi

https://ebookmeta.com/product/martingale-methods-in-statistics-
chapman-hall-crc-monographs-on-statistics-and-applied-
probability-1st-edition-nishiyama-yoichi/
Epistemic processes a basis for statistics and quantum
theory Second Edition Inge S. Helland

https://ebookmeta.com/product/epistemic-processes-a-basis-for-
statistics-and-quantum-theory-second-edition-inge-s-helland/

Research Methods and Statistics in Psychology 7th

Edition Hugh Coolican

https://ebookmeta.com/product/research-methods-and-statistics-in-
psychology-7th-edition-hugh-coolican/

Statistics and Probability Faulkner

https://ebookmeta.com/product/statistics-and-probability-
faulkner/

Dirichlet Forms and Symmetric Markov Processes

Masatoshi Fukushima Yoichi Oshima Masayoshi Takeda

https://ebookmeta.com/product/dirichlet-forms-and-symmetric-
markov-processes-masatoshi-fukushima-yoichi-oshima-masayoshi-
takeda/

Think Bayes: Bayesian Statistics in Python 2nd Edition

Downey

https://ebookmeta.com/product/think-bayes-bayesian-statistics-in-
python-2nd-edition-downey/
SpringerBriefs in Statistics
JSS Research Series in Statistics

Editors-in-Chief
Naoto Kunitomo
School of Political Science and Economics, Meiji University, Chiyoda-ku,
Tokyo, Japan

Akimichi Takemura
The Center for Data Science Education and Research, Shiga University,
Hikone, Shiga, Japan

Series Editors
Genshiro Kitagawa
Meiji Institute for Advanced Study of Mathematical Sciences, Nakano-ku,
Tokyo, Japan

Shigeyuki Matsui
Graduate School of Medicine, Nagoya University, Nagoya, Aichi, Japan

Manabu Iwasaki
School of Data Science, Yokohama City University, Yokohama, Kanagawa,
Japan

Yasuhiro Omori
Graduate School of Economics, The University of Tokyo, Bunkyo-ku,
Tokyo, Japan

Masafumi Akahira
Institute of Mathematics, University of Tsukuba, Tsukuba, Ibaraki, Japan
Masanobu Taniguchi
School of Fundamental Science and Engineering, Waseda University,
Shinjuku-ku, Tokyo, Japan

Hiroe Tsubaki
The Institute of Statistical Mathematics, Tachikawa, Tokyo, Japan

Satoshi Hattori
Faculty of Medicine, Osaka University, Suita, Osaka, Japan

Kosuke Oya
School of Economics, Osaka University, Toyonaka, Osaka, Japan

SpringerBriefs present concise summaries of cutting-edge research and

practical applications across a wide spectrum of ields. Featuring
compact volumes of 50 to 125 pages, the series covers a range of
content from professional to academic. Typical topics might include:
More information about this subseries at
A timely report of state-of-the art analytical techniques
A bridge between new research results, as published in journal
articles, and a contextual literature review
A snapshot of a hot or emerging topic
An in-depth case study or clinical example
A presentation of core concepts that students must understand in
order to make independent contributions
http://www.springer.com/series/13497
SpringerBriefs in Statistics showcase emerging theory, empirical
research, and practical application in Statistics from a global author
community.
SpringerBriefs are characterized by fast, global electronic
dissemination, standard publishing contracts, standardized manuscript
preparation and formatting guidelines, and expedited production
schedules.
Editors
Nobuaki Hoshino, Shuhei Mano and Takaaki Shimura

Pioneering Works on Distribution

Theory
In Honor of Masaaki Sibuya
1st ed. 2020
Editors
Nobuaki Hoshino
School of Economics, Kanazawa University, Kanazawa, Ishikawa, Japan

Shuhei Mano
The Institute of Statistical Mathematics, Tachikawa, Tokyo, Japan

Takaaki Shimura
The Institute of Statistical Mathematics, Tachikawa, Tokyo, Japan

ISSN 2191-544X e-ISSN 2191-5458

SpringerBriefs in Statistics
ISSN 2364-0057 e-ISSN 2364-0065
JSS Research Series in Statistics
ISBN 978-981-15-9662-9 e-ISBN 978-981-15-9663-6
https://doi.org/10.1007/978-981-15-9663-6

© The Editor(s) (if applicable) and The Author(s), under exclusive

license to Springer Nature Singapore Pte Ltd. 2020

This work is subject to copyright. All rights are solely and exclusively
licensed by the Publisher, whether the whole or part of the material is
concerned, speci ically the rights of translation, reprinting, reuse of
illustrations, recitation, broadcasting, reproduction on micro ilms or in
any other physical way, and transmission or information storage and
retrieval, electronic adaptation, computer software, or by similar or
dissimilar methodology now known or hereafter developed.

The use of general descriptive names, registered names, trademarks,

service marks, etc. in this publication does not imply, even in the
absence of a speci ic statement, that such names are exempt from the
relevant protective laws and regulations and therefore free for general
use.
The publisher, the authors and the editors are safe to assume that the
advice and information in this book are believed to be true and accurate
at the date of publication. Neither the publisher nor the authors or the
editors give a warranty, expressed or implied, with respect to the
material contained herein or for any errors or omissions that may have
been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional af iliations.

This Springer imprint is published by the registered company Springer

Nature Singapore Pte Ltd.
The registered company address is: 152 Beach Road, #21-01/04
Gateway East, Singapore 189721, Singapore
Preface
Statistical distribution theory is a traditional ield. Hence, it might be
surprising that its frontier is still lively. The present work compiled
refereed articles speci ic to this ield and showcases pioneering efforts.
Although issues discussed can be unfamiliar to even statisticians,
simple examples provided are often effective in helping readers
understand where the frontier exists.
A non-negligible part of the tradition has resulted from Professor
Masaaki Sibuya’s contributions since the 1960s. Among his various
interests in the ield, the theory of random partitioning of a positive
integer carries signi icant weight. The Ewens sampling formula is the
“normal” distribution over the partitions of a positive integer. Along
with its generalization, which is called the Pitman sampling formula,
these distributions constitute a basis for understanding the
combinatorial phenomena of random partitioning.
The present volume commences with Professor Sibuya’s latest
research on a statistical test to discriminate the Ewens sampling
formula from the Pitman sampling formula (Chapter 1). He notes the
fact that the tails of the lattice structure of partitions are more likely to
be observed under the generalized domain of the Pitman sampling
formula compared to the original Ewens sampling formula. These tails
are de ined in terms of a partial order, and fundamental thoughts on the
treatment of partially ordered sets are presented. In fact, few studies on
the use of partially ordered sets in statistics exist. We are pleased to see
Professor Sibuya’s pioneering work again in this honorary volume
dedicated to him.
The next two chapters also discuss the Ewens sampling formula.
The number of parts in the partition of a positive integer is called a
length. Professor Yamato has been working on the discrete asymptotics
of lengths as well as its approximation under the Ewens sampling
formula, as presented in Chapter 2. He provides new proof of the
convergence of length to the shifted Poisson distribution. Professor
Tsukuda employed different regimes on the asymptotics of the length
that the central limit theorem dictates. He considers two types of
standardization of length in Chapter 3. The derived results include the
error bounds of normal approximation to these types and a decay rate
depending on the growth rate of the parameter of the Ewens sampling
formula.
The partition of a positive integer induces a different story. Suppose
that a partition of a positive integer is given. Then, the ordering of its
parts generates a sequence of positive integers. In this sequence, if the
adjacent integers are the same, then this pair is called level. The
number of levels in the sequence obviously depends on the method of
ordering, but permutations exhaust all cases. A point of interest raised
by Professor Fu in Chapter 4 is the number of permutations with a
speci ic number of levels for a given partition. This issue dates back to
1755 and has been unsolved, but he has derived an explicit formula via
inite Markov chain imbedding, separating from the ordinary
combinatorial approach.
The partition also appears as the index of orthogonal polynomials.
The orthogonal property of the system of orthogonal polynomials
depends on a weight function, which is regarded as a statistical
distribution, such as the normal, gamma, or beta distributions. These
distributions can be of symmetric matrix variates, e.g., the Wishart
distribution, and Professor Chikuse is interested in orthogonal
polynomials associated with these cases. Despite the complex structure
of the orthogonal polynomials with a symmetric matrix argument, she
derived recurrence relations, which can be seen in Chapter 5.
Finally, Chapter 6 considers the parameter estimations of the
regular exponential family. Professors Yanagimoto and Ohnishi
rigorously proved that the posterior mean of the canonical parameter
vector under Jeffreys’ prior is asymptotically equivalent to the
maximum likelihood estimator, which characterizes Jeffreys’ prior. This
result itself is a mathematical elaboration rather than a pioneering
work. However, they were motivated to understand a predictor that is
expressed as a function of the posterior mean of the canonical
parameter in reference to the asymptotic behavior of the maximum
likelihood estimator. They present a step toward their vision.
These contributive articles are based on presentations at the
Pioneering Workshop on Extreme Value and Distribution Theories in
Honor of Professor Masaaki Sibuya held at the Institute of Statistical
Mathematics (ISM), Tokyo, Japan from March 21–23, 2019. This
workshop was sponsored by Kanazawa University and the ISM and co-
sponsored by the Advanced Innovation powered by the Mathematics
Platform and Japanese Society of Applied Statistics. This workshop and
the editing process for this volume were inancially supported by the
KAKENHI grant (18H00835). The editors very much appreciate the
generous support from these sponsors. Also their thanks are due to
people who cooperated in bringing this work to fruition. Particularly,
the reliability and accuracy of the information herein were made
possible due to the generosity of many esteemed anonymous
reviewers.
Last but not least, the editors would like to express their sincere
gratitude to Professor Sibuya for his decades of mentorship to them.
Nobuaki Hoshino
Shuhei Mano
Takaaki Shimura
Tokyo, Japan
August 2020
Contents
1 Gibbs Base Random Partitions
Masaaki Sibuya
2 Asymptotic and Approximate Discrete Distributions for the
Length of the Ewens Sampling Formula
Hajime Yamato
3 Error Bounds for the Normal Approximation to the Length of a
Ewens Partition
Koji Tsukuda
4 Distribution of Number of Levels in an -Speci ied Random
Permutation
James C. Fu
5 Properties of General Systems of Orthogonal Polynomials with a
Symmetric Matrix Argument
Yasuko Chikuse
6 A Characterization of Jeffreys’ Prior with Its Implications to
Likelihood Inference
Takemi Yanagimoto and Toshio Ohnishi
© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2020
N. Hoshino et al. (eds.), Pioneering Works on Distribution Theory, SpringerBriefs in
Statistics
https://doi.org/10.1007/978-981-15-9663-6_1

1. Gibbs Base Random Partitions

Masaaki Sibuya1
(1) Keio University, 3-14-1 Hiyoshi, Kohoku-ku, Yokohama-
shi Kanagawa, 223-8522, Japan

Masaaki Sibuya
Email: [email protected]

Abstract
As a typical family of random partitions on , the set of partitions of
n into k parts, the conditional distribution of Pitman’s random partition,
termed as the Gibbs base random partition, GBRP , is investigated.
The set is a lattice with respect to majorization partial order with
unique minimum and maximum, and GBRP has TP2 with respect to
this order. The main purpose of this paper is to study such a family of
random partitions and the inference on its parameter.

Keywords A-hypergeometric distributions – Distribution functions on

a majorization order poset – Ewens-Pitman sampling formula – Part-
block representation of partitions – Random partitions of n into k parts
– Total order using TP2

1.1 Introduction
The Ewens-Pitman sampling formula, EPSF , recalled in §1.2.1, is
now the representative of parametric families of random partitions,
which is applied successfully to the wide range of ields. Its special case,
the Ewens sampling formula, EPSF , is its origin and characterized
in many ways, because of speci ic features. The parameter space of
EPSF is divided into three regions, (i) the main region,
; (ii) the degenerate region,
; (iii) the Ewens sampling formula,
, the border between (i) and (ii). Hence, it is natural to question
whether an observed partition comes from the Ewens sampling
formula or not. This is the main motivation of the present paper. For
this purpose, note that given the number of parts, the conditional
partition is independent of . We refer the conditional partition as the
Gibbs base random partition, GBRP , which is formally de ined in
§1.2.3. To obtain the better understanding of EPSF, the study of GBRP
may be helpful as a complement of EPSF . This is the other
motivation for the present research.
The set of all partitions of n into k parts, , or its union
is partially ordered by majorization, and they are
lattices with unique maximum and minimum partitions. Along with the
majorization order, the probability mass function (pmf) of partitions of
GBRP is TP2, and it changes smoothly on . Unexpectedly,
is not a special parameter value in GBRP except for that some
expressions are simple at .
Concerning EPSF, the recent survey by Crane [3] with discussions is
highly recommended. See also text books by Feng [5] and Pitman [15]
and the preceding survey by Tavaré and Ewens [21]. Random partitions
on are called microcanonical distributions in physics, and the
speci ic role of GBRP within the Gibbs random partitions is shown by
Gnedin and Pitman [8].
 In the statistical theory, the concept of partial order is not popular.
Among some of the works dealing with partially ordered sets (posets), a
typical subject is the measure of agreement between two types of
posets; see [16]. Another typical subject is the concept of partial
quantile de ining the quantile of the random variable in the general
poset; see [2]. Since we are concerned with the parametric family of
probability distributions on inite posets, our approach is different from
theirs.
Although the current topic on random partitions is Bayesian
nonparametric statistics (see, e.g., [6] or [7]), the present paper is
classical and elementary, and numerical calculations and visualizations
are devised. A more profound study, regarding GBRP as an A-
hypergeometric distribution de ined by Takayama et al. [20] was done
by Mano [13].
In the remaining part of introduction, the partition of a number and
the majorization order of partitions are reviewed in short. In Sect. 1.2,
the properties of GBRP , including the parameter estimation related
to [13], are studied. In Sect. 1.3 the tests of hypotheses on are
studied, and problems in numerical computations are discussed. In
Appendix, lower adjacency in GBRP, the majorization order of integer
vectors, and partial order in GBRP are brie ly discussed.

1.1.1 Partitions of a Number

Let denote the set of positive integers. If and
, the set is a
partition of n, expressed as . The elements ’s are called parts
(clusters, blocks, or species). If , is equivalent
to . Its three expressions are speci ied as follows. The irst
expression is descending order statistics (dos):
Its visualization, the Ferrers diagram (or the Young diagram), is a layout
of n squares in k lush left rows with lengths . The second
expression is the size index (the terminology proposed by Sibuya [18]
for the size of parts):

The third one is a combination of the above two or the list of part-
blocks:

which is the reversed sequence of the classical expression

and usual observation records.
For example, the three expressions of an example partition, and a
corresponding Ferrers diagram with part-block frames are shown
below.

These three expressions will be liberally used to denote a partition.

Hereinafter, letters n, k, and b are used exclusively in the above sense
throughout the present paper. The set of partitions of n into k parts is
denoted by . Further, , . and
are called partition numbers.
Proposition 1.1 The minimum number of part-blocks in is
or 2. If , there is a partition of one block: .

Otherwise, there are a number of partitions of .

The maximum number of part-blocks in is as follows.
If , then . Otherwise, is the solution of

Proof (of the maximum part) It is easy to see, iff .

Discard the leftmost column of the Ferrers diagram, that is, subtract
one from all parts. Then (squares) must be allocated into
parts. The above inequlity, being applied to the pair
, implies that, if , then .

The minimum satisfying , say , provides

. In other words, .

Figure 1.1 illustrates the maximum number of part-blocks of partitions

in . The maximum number in is approximately .

Proposition 1.1 is intended to check the ef iciency of part-block-wise

operations compared to part-wise operations when n and k are large.
An example is the function genpart(n, k, m) given in the next paragraph.
The depth of recursive calls is substantially reduced.

Fig. 1.1 Maximum number of part-blocks of partitions in

The partitions of are totally (or linearly) ordered by the reverse

lexicographical order. For example, if , in dos expression,

In this paper, the priority is given to the order of k, and within ,

partitions are ordered by the reverse lexicographical order (rlg order).
For example, in ,

(1.1)
A way for generating all partitions of is the recursive call of the
function genpart(n, k, m), with . The function generates
all partitions of whose maximum size of parts is restricted by m.
The result is ordered like (1.1). A useful total order, which is speci ic to
GBRP, is introduced in §1.3.1.

A well-known partial order for , ,

is de ined by

Note that implies

Geometrically, the Ferrers diagram of is included in that of .
For any two elements and of the partially ordered set (POS)
, two binary operators and are de ined as follows,
regarding them as the in inite sequence with zeros tail.
The operator denotes union, and denotes meet, of the Ferrers
diagrams. For any , there exist uniquely and .
Hence is a lattice, and called the Young lattice.

1.1.2 Majorization Order of Partitions

The Young lattice is strong and orders partitions of different n, whereas
we are interested in weaker order for comparison within or .
The majorization order (or dominance order) for
and in is de ined by

Example
. The layout is increasing rightward and upward.

Note that and . This example is too simple, and the

layout is misleading. For larger n, even restricted to , partitions

cannot be arranged on a plane grid.

Binary operators and are de ined within , that is;

Proposition 1.2 For any , there exist unique and
within . Hence, is a lattice.

Proof If or , the proposition is trivial. In general,

it is proven constructively. Let the cumulative sum of dos expressions
be denoted as , , , then and

are increasing and concave functions in j, and

, is also increasing and concave. Hence,
is positive, decreasing, , and .
Next, let , , and
is not always decreasing, so let d be the
rearrangement of in decreasing order. Since d is the minimum of
concave majorants of and , .

Since , , the cumulative

sum of dos, is another expression of partitions.
There exist unique lower and upper limits in :

Note that both limits have two or one part-blocks.

Proposition 1.2 is extended to the fact that is a lattice.
Further, in ,
and are lower and upper limits, respectively. is its
special case where and are limits.
The theory of the majorization order is an important and
established principle. All aspects of this theory are covered in Marshall
et al. [14]. See a paragraph in Appendix.

Proposition 1.3 The majorization order implies both

lexicographical order and reverse lexicographical order (rlg), ,
(1.1), in and .

Proof If , , which means that for

some , . Hence .

1.1.3 Adjacency
To see iner structures in POS , we de ine a relation , is
smaller than and adjacent to (or ), if

For examining the majorization order structure in (or ), we

need to identify the upper or larger adjacent partitions of each partition
of (or ). For this purpose, make a part-block less equal; take an
element of a part-block, and move it to another part of the same part-
block, otherwise, if the part-block consists of a single part, move it to a
larger part. According to the choice of a part-block, the change can be
one of the following two moves:
(A)
Move to an upper partition of .
(B)
Move from the partitions of to a partition of .

We can consider a dual way: (C) Move to a lower partition of ,

and (D) Move from the partitions of to a partition of . These
moves are summarized in Table 1.10 in Appendix.
Further details are as follows. The remarked part-block will be
denoted by .
A1. Condition and .
Take an element of a part-block such that and
, and put it into another part of the same part-block. The result of this
move is classi ied further as follows:
a.
If and ,

b.
If and ,

c.
If and ,

In these results, if , the middle part-block disappears. The

symbol denotes unchanged part-blocks, and denotes generic
part-blocks affected by as explained later.
In A1, the case of and is excluded. In this situation,
move an element of to the upper block , and the
result will be

However, this result can be obtained through two moves. First, apply
move A1 to to obtain
Second, move an element of to the upper part-block. Hence, the
above move does not result in adjacent partition, provided that
, which is a necessary condition in the above two move
argument. Hence, the last possible situation is
A2. Condition and .
In this speci ic situation, an element of can be moved into
.

a.
If , the irst two part-blocks are involved, and the result is

b.
Generally, if and , two part-blocks are involved,
and the result is similar to the case a.

c.
Finally, if and , three part-blocks are involved,
and the result is

Throughout A1 and A2, the last part-block and should be

remarked. If is the last part-block, namely, , the new last
part-block is . If , the succeeding part-blocks are
unchanged, and if , these will be .
Proposition 1.4 The change of a partition in to the adjacent
upper partitions in , by the move of an element of a part-block, is
summarized in part A of Table 1.2.

See, as an example, Table 1.1 of adjacency in .

Table 1.1 Adjacency in . There are 52 adjacent pairs. Partitions are

arranged from the lower to the upper. In 1(136) etc., the parenthesized is the rlg
order in

Let , be the size index expression of . The

result of Move A1 and A2 of is simple in this expression:

(1.2)

(1.3)
If , which is the case excluded from case A, a part of the last part-
block disappears, and the partition changes from to . The
details are similar to case A. Remark that means that the part-
block is the last one, so that , that is, the number of part-blocks of
the partition.
B1. Condition and .
Take an element of the part-block such that , and put
it into another part of the same part-block.
a.
If , the partition has only one part-block that and
changes to
b.
If and , the last part-block changes to

c.
If and , it follows that

Table 1.2 Classi ication of the upward transforms from a partition

. The type of part-blocks, from which one element is to be taken
out, is classi ied. A: within , B: from to
 B2. Condition and .
Note that if , as in Case A2, adjacent upper partitions are
possible only if . The last part-block is (1, 1), and its element is
moved to the upper adjacent part-blocks.
a.
If and , the new partition is

b.
If and , the new last part-block is

c. If and , the new last part-block is

Proposition 1.5 The change of a partition in to the adjacent
upper partitions in , by the move of an element of a part-block, is
summarized in part B of Table 1.2.

In the size index expression of , the results of Move B1 and B2 of are

as follows:

Its implication is out of the scope of the present paper.

Upper set generation
Given , we construct its upper set
. All partitions will be expressed in dos, or equivalently, in a cumulative
sum of dos sequence. Let denote the maximum partition of
, or of . That is, , and

where , and represents the

component-wise comparison. The cumulative sum of a partition of

is shown as an increasing concave polygonal line limited by those
of and . Figure 1.2 illustrates of of an
example that , where and in .

See also Table 1.1 of adjacency in .

Fig. 1.2 Cumulative sum expression of partitions of

First, we describe the outline of the algorithm for generating an

upper set. In Fig. 1.2, all lines connecting and are
such that . From these candidates, we select
the increasing and concave lines:

The problem here is how to avoid redundant works and save memory.
The output of the algorithm is a matrix, whose columns are
cumulative sums of , combined in rlg order. We construct
the matrix starting from its bottom, matrix [n], moving up row by
row and thereby, increasing the number of columns.
Algorithm for . If has a part-block , the last c segments
of broken lines are reduced to one straight line:
 . We start from a matrix :

The procedure for constructing from is as

follows.
The candidates to append to subsequences (columns) are
. Append the vector of to
as the top row; check the new columns whether they are

strictly increasing and concave, and delete inadequate columns. Thus,

the columns of are reduced to, say, . Next, append the vector

of to as the top row, and delete inadequate columns.

Repeat this procedure up to , and join the matrices column-

wise to obtain .
If is obtained, the inal procedure is to append the row vector of
zeros on the top of , and check the concavity. Reduced is the
cumulative sum expression of .
Moving to the next , the matrix is reduced at

each step for ef iciency. If is decreasing or convex,

 cannot be strictly increasing or concave.
A numerical example for of Fig. 1.2 is provided as follows:

The construction of lower sets is a conjugate procedure. If

, , and the procedure is another way for
constructing , with the duplicate adjacency information .
Alternative af irmation of adjacency
The relation can be easily checked, and the other way to
af irm adjacency in should be considered. Let
, be a sequence of partitions in rlg order. Let

M be a logical matrix with the (i, j) element

Hence, T (true) elements are restricted to the lower triangle of M, and

the diagonal elements are trivially T.
 Here, the purpose is to construct the logical matrix with the
(i, j) element

from M. The procedure goes downwards and by rows. In the i-th row of
M, is the upper set of , and from

these partitions, delete upper partitions of recorded temporarily in

, the j-th row of . Going through , is

determined accordingly. Formally, the procedure is described below:

Algorithm.

for

set ,
for
if , for all such as , set
set

1.2 Gibbs Base Random Partitions

1.2.1 The Ewens-Pitman Sampling Formula
The Ewens-Pitman sampling formula, EPSF , or simply, the Pitman
random partition is de ined by

(1.4)

where
and s is the size index expression of a partition. The generalized
factorial product (x|c) includes the usual descending factorial product
, the ascending factorial product , and the power
. Its parameter space is

On the boundary that , , and on the

boundary that , . In the following, these
degenerated random partitions will be neglected.
A special case,

is the Ewens sampling formula. Another special case,

is the symmetric negative hypergeometric distribution.

The number of parts
Let be the size index of a random partition in .
is the number of parts, and ,
referred to as EPSF-K , satis ies the forward equation:
and

(1.5)

where is the exponential generalized Stirling numbers; see,

e.g., [19]. See also the discussion following the de inition of GBRP in
§1.2.3.
For the Ewens sampling formula, K is a suf icient statistics, and

Conditional distribution on
The conditional probability,

the ratio of (1.4) to (1.5), is independent of and equals to (1.6), the

Gibbs base random partition de ined in §1.2.3.

1.2.2 Probability Measure on

Stochastic orders on lattice
Three elemental stochastic orders, usual, hazard rate (reverse
hazard rate), and likelihood ratio, and their implications are well
known. See, e.g., [17]. Here, they are described in terms of the discrete
partially ordered set with the lattice structure, with unique
minimum and maximum. Let , and denote the distribution
function (df), the survival function (sf), and the pmf, respectively, of a
random quantity X on a lattice . Those of Y are similarly denoted.
The usual stochastic order, , is de ined by

The hazard rate order, , is de ined by

Its dual order, reverse hazard rate order, , is de ined by

The likelihood ratio order, , is de ined by

( by convention). That is,

The order implies and , and each of these

implies . These comparisons are applied to the monotonicity of
a parametric family of random partitions.
Distribution function and survival function on
Let , a lower set of , be the set of the partitions of that
are lower than . The upper set is similarly de ined:

A probability measure P on is de ined by a probability mass

function , and
are the distribution function and the survival function, respectively, of
or . The relationship between them is

As a special case of the previous paragraph, implies that

and are nondecreasing.
An example is illustrated in §1.2.3. The notions can be extended to a
partial order set in general.

1.2.3 Gibbs Base Random Partitions

The Gibbs base random partition is the conditional random partition of
the Ewens-Pitman sampling formula, EPSF , on . Using size
index expression,

(1.6)

See (1.4) for the notations. In particular,

Note that , a special case of EPSF , is a random

partition with parameter only. Hence, its relation with GBRP is
concerned.
Hence

(1.7)

This polynomial identity in is another de inition of .

Moments of the size index
Let be the size index of GBRP on . The
joint factorial moments are

(1.8)

where

Especially,
Note that, if , the r.h.s. of is not but .

1.3 Testing Statistical Hypotheses on

1.3.1 Properties of GBRP
First, we investigate the pmf of GBRP as a function of .

Proposition 1.6 The pmf of the maximum ( ) partition of is

increasing in , while that of the minimum ( ) is decreasing. Such a
partition increasing near is unique in . If there is another
partition with a decreasing pmf, those in its lower set are decreasing. In
this sense, there is a lower set with decreasing pmfs. Hence, the
maximum likelihood estimate does not exist for these tail partitions. It
is anticipated that the pmf of a partition between ( ) the two tail sets
is unimodal.

Proof First, the largest partition is shown to be increasing. Denote the

pmf of a partition as . For any ,
. This is a polynomial of with
degree ; the same for all . Since
 is strictly increasing in , and
, , is strictly increasing

. Hence,

is decreasing. Similarly, the minimum partition of is shown

decreasing.
Next, such a partition as is unique in .
Because of (1.5) or (1.6), is a polynomial in of degree
with single root , and the other real roots are larger than 1.
Moreover, has the factor (size index expression),

and

The last condition is satis ied only by .

If the pmf of a partition is decreasing, according to Theorem 1.1

below, it is straightforward that partitions smaller than it has a
decreasing pmf. The existence of a partition with a monotone pmf other
than or and the unimodality of other pmfs are open problems.

The fact that the mle does not exist at the tails of is shown
asymptotically by Mano (see Fig. 5.3 of [13]). Proposition 1.6 is an
elemental justi ication of his result.
Table 1.3 provides a small numerical example of the likelihood and
monotone tail partitions. In the "likelihood" rows, and mean that
likelihood is decreasing or increasing, respectively, while the numbers
represent the maximum likelihood estimates.
Table 1.3 Monotone and unimodal pmf of . In the rows of mle,
and show that the likelihood is decreasing and increasing, respectively

The majorization order of partitions in is as follows. The

partitions are increasing upward and rightward:

The pmf of the partitions of is illustrated in Fig. 1.3, and the

corresponding sf and df are illustrated in Fig. 1.4. The abscissa of these
igures is in a total order , which will be introduced in Paragraph
Total order and visualization. The ordinate of Fig. 1.3 is in reverse
lexicographical order, - rlg.
Fig. 1.3 Probability mass functions of GBRP on . The

area of gray squares is proportional to a pmf. Segments show the adjacency of

partitions. The leftmost lowest square represents , and the rightmost highest

one does . Upper: ; Lower:

Fig. 1.4 Distribution functions and survival functions of GBRP, . The top

line: df for ; the bottom line: df for and sf for .

Vertical broken lines show

Let denote the pmf of GBRP.

Theorem 1.1 The pmf , is totally positive with

degree 2, TP2:

(1.9)

is increasing in .
Proof It is enough to show that

is increasing in . Such a pair is speci ied, by

Proposition 1.4, to those that are associated with Moves A1 and A2. In
§1.1.3, the change of a partition by Moves A1 and A2 is expressed in the
size index: (1.2) and (1.3). That is, the size index changes,
by Move A1, to .
Hence, for Move A1,

For Move A2, the result is

, and

That is, is totally positive of order 2 in the majorization order sense.

In other words, is Schur-convex in terms of the majorization theory.
The related implications will be discussed in the next section.
Total orders and visualization
Theorem 1.1 shows that , a function in depending on
arbitrarily chosen , de ines a total (linear) order of
, which is more useful than the reverse lexicographical
order for dealing with GBRP. Here, letter a is used instead of , because
this paragraph is apart from inference. By changing , the order
of incomparable partitions changes, but its effect is not clear. We
propose the simplest one in Proposition 1.7.
A standard selection of is as follows:

Proposition 1.7

is increasing in .

Proof From Theorem 1.1,

is increasing in . is its special case .

Figure 1.5 shows the region of (a0, a1), where the total order by
is the same as that by for . The size of the
region is smaller for larger n, and the shape in larger a1 depends on k.
Fig. 1.5 (a0, a1) where the orders by and by are the same for

. Lines are of , and

Hence, is a standard way of total ordering . The

abscissa of Figs. 1.3 and 1.4 is of . The total ordering of the
partitions of GBRP ( ) by will be denoted by std.
Standardization of
The range of changes by different (n, k), and standardization

stabilizes the distribution of . To illustrate the effect of

standardization, the nine patterns of dos or the part-blocks of different
(n, k) are selected, and for each partition, the probability of the upper
set versus is plotted in Fig. 1.6.
 The nine patterns, in S/R expression, are as follows:

The abscissa of Fig. 1.6 is of .

Fig. 1.6 Probability of the upper set versus in several patterns of partitions

1.3.2 Testing Statistical Hypotheses

One-sided test
One of the motivations of this paper is to test the hypothesis that
. An immediate answer is the likelihood ratio test, based on
Theorem 1.1.
 In the one-parameter family of univariate continuous distributions,
a distribution belongs to a monotone likelihood ratio family if
the ratio is nondecreasing in a function T(x). In
Theorem 1.1, the ratio depends on , and the uniformly most
powerful test does not exist.
In GBRP on , for testing

let be an upper set of a partition . We observe four facts:

1.
is the rejection region of the test with the level .
2.
The power of the test, , is increasing in . That is, the
test is unbiased.
3.
Randomization test is possible for controlling the level; however,
for small n and k, the controllability is limited.
4.
The uniformly most powerful test does not exist.
See [11] for the monotone likelihood family.
There are more than one upper sets that have close levels. For
selecting one among them, , a possibility is to select

such as is large. Let and be candidates; that is,

De ine
Another random document with
no related content on Scribd:
A Penny-in-the-Slot Newspaper Stand
 A silent, yet very successful “newsboy” is shown in the photograph
reproduced. Each rack holds one of the daily newspapers. The
passer-by puts a penny in the slot and selects his paper. The penny
falls down a chute, striking a bell, as evidence that the customer has
paid. Hundreds of papers are sold from this rack each day without
the aid of the newsboy’s shouts. The slanting shelves make it
unnecessary to provide weights for the papers.—M. F. Hall, St.
Louis, Mo.
Pedals for Typewriter Space and Shift Keys
Increase Speed
 By attaching pedals to the space key and shift key I increased my
typewriting speed about 20 words a minute. I made two pedals fast
to the typewriter desk, as shown, then ran stiff steel wires up to the
keys, connecting the latter to the wire with rubber bands. The pedals
are adjusted so that a light tap operates them. If a light typewriter is
used, it should be fixed to the desk. The wires can be unhooked
quickly from the typewriter, and suspended on hooks under the edge
of the desk.—W. J. M., Manila, P. I.
 Timing Pointer on Watch Crystal
Working in a photo-print finishing plant where often several jobs
need attention and accurate timing, all at the same time, I find a
pointer on my watch crystal a great help. The crystal was loose
enough to permit turning it. I glued a three-cornered piece of black
paper to the under side of the crystal at the edge. When a batch of
prints is placed in a bath, the pointer is set 15 minutes ahead of the
minute hand. When the latter reaches the pointer, it reminds me to
remove the prints. After placing a batch of films in the developing
tank, I set the pointer ahead of the minute hand the required time for
developing, placing the watch in view, to see readily when the time is
up.—A. J. Viken, Waterloo, Iowa.
 Portable Tent Made from an Umbrella and Paper
Muslin

Picnickers desiring to go in bathing are often handicapped by the

lack of a convenient place to change clothes. An umbrella and some
paper muslin provides a light portable tent that is practical and
inexpensive for such uses. Cut the dark paper muslin into as many
9-ft. lengths as there are sections of the umbrella. Sew these strips
together. At each seam tie a string about a yard long and a stout
cord 15 ft. long to the handle to hold up the tent. For use, open the
umbrella, invert it, and to each rib tie one of the strings. Then tie one
end of the cord to the handle of the umbrella and suspend it from a
tree or other support as shown, weighting or tying down the other
end.—Mabel Hubbard Birch, Cincinnati, Ohio.
Memorandum List for Camping and Outing Trips
Persons who occasionally take an outing or go camping will find a
written list of articles needed on the trip a great saving of time and
mental energy. This list should be made while making preparations,
and again read shortly before the departure. It may be kept from
year to year, and items added or withdrawn, as experience teaches.
Classification is desirable, under heads such as camping gear,
cooking utensils, tableware, foods, supplies, etc. Each member of a
party can agree to furnish certain articles, avoiding duplication.
Anyone who has gone on outings and left behind some much-
needed article, will readily realize the help such a list can be, and by
its use the amateur may profit from the experience of others.
 A Homemade Trouble Lamp

By attaching a handle for carrying, and making suitable

connections, a portable trouble light was made of a dry battery. A
strip of iron was bent to the shape shown, for the handle, and fixed
to the binding post on the carbon. A bell from an alarm clock was
drilled and tapped to fit a small light bulb, and fastened at the zinc
post. The upper edge of the bell was bolted to the extension of the
handle, as shown, leather washers insulating it. A small switch,
riveted loosely to the iron strip, controls the light.—Carl A. Haberlein,
McPherson, Kan.
¶Books to be stacked or carried should be arranged with the backs
on alternate sides.
 Pulley and Weight Exerciser Homemade in the
Orient

Fig. 1 Fig. 2

When the Necessary Hardware Was Not at Hand, a Satisfactory Rigging was
Made for Exercising, by the Use of Novel Substitute Materials

Finding it difficult to obtain materials for the making of a gymnastic

exerciser, I made one with makeshift parts, typical of this locality.
Later I was able to get the necessary hardware and made an
exerciser, as shown in Fig. 1, quite similar to those manufactured
commercially. The original arrangement is shown in Fig. 2. I mounted
a bamboo rod on two iron brackets fastened to the wall. Small pegs
provided guides for the ropes, and the smooth surface of the
bamboo caused little wear. Other bamboo rods were used for the
handle and to support the weights. These were made by filling cans
with sand, and tying them with cords. This arrangement was quite
practical, but did not give the desired freedom of movement.—M. A.
Laner, Hongkong, China.
 Scoop on Painter’s Knife Catches Scrapings
By fitting a sheet-metal scoop to a painter’s putty knife or wall
scraper, much of the dirt and refuse resulting is caught in the
receptacle instead of falling to the floor and being tracked into other
rooms. The scoop is a rectangular can, and is fastened into place by
small bolts fitted through the scraper.
 Gas-Hose Bracket for Ironing or Shop Use

The Bracket Gives Greater Freedom in the Use of the Gas Iron

In a home where it was customary to use an overhead gas fixture

for the attachment of the gas iron for ironing clothes, it was found
that mantles were frequently broken by the resulting jar. By making a
bracket for the gas hose, which is fastened to the end of the ironing
board or to the edge of the table, this trouble was remedied. The
new arrangement gave much greater freedom in the use of the iron.
The bracket can be removed quickly and fitted to a bench for use in
the shop.—M. J. F. Schramm, Ridgewood, N. Y.
 Homemade Guide for Photo and Paper Trimmer

The Spring Hinge Presses the Guide Down on the Print being Cut

Having an old but still serviceable photograph trimmer, with which

it was difficult to trim prints uniformly and quickly, I made a guide for
this purpose and fitted it to the board, as shown. The guide is a strip
of wood, ¹⁄₄ in. thick, and is fastened to the cleat with a spring hinge,
mounted on a block. Although it is not transparent, like the best
commercial devices of this kind, it gives good service.—R. B. Manck,
Nazareth, Pa.
 Joining Broken Liquid Column in Thermometer
To reunite the particles of mercury or other liquid, which have
become separated in a thermometer, raise it to the level of the
extended arm, the bulb outward, and jolt the tube with a quick
downward stroke. Repeat this, not too vigorously, until the liquid
returns to its proper position.
 Umbrella Used as a Clothes Drier

A clothes drier that can be carried in a purse or a vest pocket will

appeal to travelers, and persons living in small quarters. An
umbrella, four yards of strong wrapping twine, and several small
brass rings are required. Knot the rings into the twine at intervals,
measuring the distance between the rib points of the umbrella, and
hook the twine to the points by the rings, as shown, providing
considerable drying space for small articles. Hook the umbrella
handle over a suitable support, or tie it carefully to the supporting
pipe of a light fixture in the middle of the room, ready for the articles
to be dried. The twine may also be wrapped around the points, but it
is better to carry a line with rings attached.

¶By reversing the handle on a dust brush which has been worn near
the point, additional wear may be obtained.