Partitions Hypergeometric Systems and Dirichlet Processes in Statistics Springerbriefs in Statistics Mano Shuhei
Partitions Hypergeometric Systems and Dirichlet Processes in Statistics Springerbriefs in Statistics Mano Shuhei
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/
https://ebookmeta.com/product/introduction-to-mathematical-
statistics-whats-new-in-statistics-8th-edition-robert-hogg/
https://ebookmeta.com/product/basic-statistics-in-business-and-
economics-ise-hed-irwin-statistics-10th-edition-douglas-a-lind/
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/
https://ebookmeta.com/product/research-methods-and-statistics-in-
psychology-7th-edition-hugh-coolican/
https://ebookmeta.com/product/statistics-and-probability-
faulkner/
https://ebookmeta.com/product/dirichlet-forms-and-symmetric-
markov-processes-masatoshi-fukushima-yoichi-oshima-masayoshi-
takeda/
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
Shuhei Mano
The Institute of Statistical Mathematics, Tachikawa, Tokyo, Japan
Takaaki Shimura
The Institute of Statistical Mathematics, Tachikawa, Tokyo, Japan
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.
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.
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.
The third one is a combination of the above two or the list of part-
blocks:
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 .
. In other words, .
(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.
Example
. The layout is increasing rightward and upward.
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
b.
If and ,
c.
If and ,
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
(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
b.
If and , the new last part-block is
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 :
from M. The procedure goes downwards and by rows. In the i-th row of
M, is the upper set of , and from
Algorithm.
for
set ,
for
if , for all such as , set
set
(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
(1.5)
Conditional distribution on
The conditional probability,
(1.6)
(1.7)
(1.8)
where
Especially,
Note that, if , the r.h.s. of is not but .
. Hence,
and
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
(1.9)
is increasing in .
Proof It is enough to show that
Proposition 1.7
is increasing in .
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
Fig. 1.6 Probability of the upper set versus in several patterns of partitions
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
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
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
The Bracket Gives Greater Freedom in the Use of the Gas Iron
The Spring Hinge Presses the Guide Down on the Print being Cut
¶By reversing the handle on a dust brush which has been worn near
the point, additional wear may be obtained.