## Seminar on Probability and Statistics

Organizer(s) Nakahiro Yoshida, Hiroki Masuda, Teppei Ogihara, Yuta Koike

Seminar information archive

### 2010/04/28

15:00-16:10   Room #002 (Graduate School of Math. Sci. Bldg.)
KATO, Shogo (The Institute of Statistical Mathematics)
A Markov process for circular data (JAPANESE)
[ Abstract ]
We propose a discrete-time Markov process which takes values on the unit circle. Some properties of the process, including the limiting behaviour and ergodicity, are investigated. Many computations associated with this process are shown to be greatly simplified if the variables and parameters of the model are represented in terms of complex numbers. The proposed model is compared with an existing Markov process for circular data. A simulation study is made to illustrate the mathematical properties of the model. Statistical inference for the process is briefly considered.
[ Reference URL ]
https://www.ms.u-tokyo.ac.jp/~kengok/statseminar/2010/00.html

### 2010/03/29

13:00-14:10   Room #002 (Graduate School of Math. Sci. Bldg.)
Catherine Laredo (MIA, INRA)
Inference for partially observed Markov processes and applications
[ Abstract ]
We present some statistical methods for estimating the param- eters of a population dynamics model of annual plants. It is modelled using multitype branching processes with immigration. The data consist of counts in each type that are measured in several populations for a few consecu- tive years. Parametric inference is first carried out when count data of all types are observed. We prove statistical identifiability for all the parameters ruling the population dynamics model and derive consistent and asymptot- ically Gaussian estimators. However, it often occurs that, in practice, one or more types cannot be observed, leading to partially observed processes. Parametric inference is first studied in the case of Poisson distributions. We characterize the parameter subset where identifiability holds and de- rive consistent and asymptotically normal estimators for this parameter subset. Theses results are then extended to other distributions.

We apply these results to feral oilseed data. The model takes account of reproduction, immigration, and seed survival in a seed bank. The data consist of the number of plants in several developmental stages that were measured in a number of populations for few consecutive years. They are incomplete since seeds could not be counted.
[ Reference URL ]
https://www.ms.u-tokyo.ac.jp/~kengok/statseminar/2009/17.html

### 2010/03/15

15:00-16:00   Room #002 (Graduate School of Math. Sci. Bldg.)
Cecilia Mancini (University of Florence)
BROWNIAN COVARIATION AND CO-JUMPS, GIVEN DISCRETE OBSERVATION
[ Abstract ]
We consider two processes driven by Brownian motions plus drift and possibly infinite activity jumps.

Given discrete observations we separately estimate the covariation between the two Brownian parts and the sum of the co-jumps. This allows to gain insight into the dependence structure of the processes and has important applications in finance.

Our estimator is based on a threshold principle allowing to isolate the jumps over a threshold.

The estimator of the continuous covariation is asymptotically Gaussian and converges at speed square root of n when the jump components have finite variation. In presence infinite variation jumps the speed is heavily influenced both by the small jumps dependence structure and by their jump activity indexes.

This talk is based on Mancini and Gobbi (2009), and Mancini (2010).
[ Reference URL ]
https://www.ms.u-tokyo.ac.jp/~kengok/statseminar/2009/16.html

### 2010/03/15

14:00-15:00   Room #002 (Graduate School of Math. Sci. Bldg.)
Alexandre Brouste (Université du Maine)
Statistical inference in the partial observation setting, in continuous time
[ Abstract ]
In various fields, the {\\it signal} process, whose law depends on an unknown parameter $artheta \\in \\Theta \\subset \\R^p$, can not be observed directly but only through an {\\it observation} process. We will talk about the so called fractional partial observation setting, where the observation process $Y=\\left( Y_t, t \\geq 0 ight)$ is given by a stochastic differential equation: egin{equation} \\label{mod:modelgeneral} Y_t = Y_0 + \\int_0^t h(X_s, artheta) ds + \\sigma W^H_t\\,, \\quad t > 0 \ where the function $h: \\, \\R imes \\Theta \\longrightarrow \\R$ and the constant $\\sigma>0$ are known and the noise $\\left( W^H_t\\,, t\\geq 0 ight)$ is a fractional Brownian motion valued in $\\R$ independent of the signal process $X$ and the initial condition $Y_0$. In this setting, the estimation of the unknown parameter $artheta \\in \\Theta$ given the observation of the continuous sample path $Y^T=\\left( Y_t , 0 \\leq t \\leq T ight)$, $T>0$, naturally arises.
[ Reference URL ]
https://www.ms.u-tokyo.ac.jp/~kengok/statseminar/2009/15.html

### 2010/02/17

15:00-16:10   Room #128 (Graduate School of Math. Sci. Bldg.)

[ Abstract ]

[ Reference URL ]
https://www.ms.u-tokyo.ac.jp/~kengok/statseminar/2009/14.html

### 2009/12/21

15:00-16:10   Room #128 (Graduate School of Math. Sci. Bldg.)
Thomas Simon (Universite de Lille 1)
Absolute continuity of Ornstein-Uhlenbeck processes
[ Abstract ]
Let X be a multidimensional Ornstein-Uhlenbeck process, solution to the S.D.E.

dX = AX + dB

where A is a real nxn matrix and B a Lévy process. We show that when A is non-singular, the law of X_1 is absolutely continuous if and only if the jumping measure of B fulfils a certain geometric condition with respect to A and the Gaussian part of B, which we call the exhaustion property. This optimal criterion is much weaker than for B, which might be very singular and genuinely one-dimensional. The proof uses a certain time derivation procedure and basic arguments from controllability theory.
[ Reference URL ]
https://www.ms.u-tokyo.ac.jp/~kengok/statseminar/2009/13.html

### 2009/12/16

15:00-16:10   Room #128 (Graduate School of Math. Sci. Bldg.)
Stefano Maria Iacus (Department of Economics, Business and Statistics, University of Milan, Italy)
ecent results on volatility change point analysis for discretely sampled stochastic differential equations
[ Abstract ]
In this seminar we review recent advances on change point analysis for the volatility component of stochastic differential equations under different discrete sampling schemes. We consider both ergodic and high frequency and non ergodic cases. Results have been obtained by means of least squares, CUSUM tests and quasi-maximum likelihood approach. We show an application to the recent financial crisis and finally present a Monte Carlo study to compare the three methods under different setups.

Join work with Prof. Nakahiro Yoshida.
[ Reference URL ]
https://www.ms.u-tokyo.ac.jp/~kengok/statseminar/2009/12.html

### 2009/12/14

14:00-15:10   Room #128 (Graduate School of Math. Sci. Bldg.)
L. VOSTRIKOVA (LAREMA, Departement de Mathematiques, Universite d’Angers, FRANCE)
On the stability of contingent claimes in incomplet models under statistical estimations.
[ Abstract ]
In exponential semi-martingale setting for risky asset we estimate the difference of prices of options when initial physical measure P and corresponding martingale measure Q change to tilde{P} and tilde{Q} respectively. Then, we estimate L1 distance of option’s prices for corresponding parametric models with known and estimated parameters. The results are applied to exponential Levy models with special choise of martingale measure as Esscher measure, minimal entropy measure and f^q -minimal martingale measure. We illustrate our results by considering GMY and CGMY models.
[ Reference URL ]
https://www.ms.u-tokyo.ac.jp/~kengok/statseminar/2009/11.html

### 2009/12/09

15:00-16:10   Room #002 (Graduate School of Math. Sci. Bldg.)

[ Abstract ]

[ Reference URL ]
https://www.ms.u-tokyo.ac.jp/~kengok/statseminar/2009/10.html

### 2009/11/27

13:40-14:50   Room #128 (Graduate School of Math. Sci. Bldg.)

[ Abstract ]
In this talk, we consider the problem of estimating the innovation density in nonlinear autoregressive models. Specifically, we establish the convergence rate of the supremum distance between the residual-based kernel density estimator and the kernel density estimator using the unobservable actual innovation variables. The proof of the main theorem relies on empirical process theory instead of the conventional Taylor expansion approach. As applications, we obtain the exact rate of weak uniform consistency on the whole line, pointwise asymptotic normality of the residual-based kernel density estimator and the asymptotic distribution of a Bickel-Rosenblatt type global measure statistic related to it. We also examine the conditions of the main theorem for some specic time series model.
[ Reference URL ]
https://www.ms.u-tokyo.ac.jp/~kengok/statseminar/2009/09.html

### 2009/10/23

15:00-16:10   Room #128 (Graduate School of Math. Sci. Bldg.)
On invariant measures of diffusion processes with unbounded drifts
[ Reference URL ]
https://www.ms.u-tokyo.ac.jp/~kengok/statseminar/2009/08.html

### 2009/10/22

16:30-17:40   Room #122 (Graduate School of Math. Sci. Bldg.)

ASYMPTOTICALLY EFFICIENT DISCRETE HEDGING
[ Abstract ]
The notion of asymptotic efficiency for discrete hedging is introduced and a discretizing strategy which is asymptotically efficient is given explicitly. A lower bound for asymptotic risk of discrete hedging is given, which is attained by a simple discretization scheme. Numerical results for delta hedging in the Black-Scholes model are also presented.
[ Reference URL ]
https://www.ms.u-tokyo.ac.jp/~kengok/statseminar/2009/07.html

### 2009/10/21

15:00-16:10   Room #002 (Graduate School of Math. Sci. Bldg.)

AR過程の優調和事前分布と偏自己相関係数による表示
[ Abstract ]
Tanaka and Komaki(2008)では時系列データが2次の自己回帰過程(AR過程)に従う 時のスペクトル密度の推定を考え、優調和事前分布に基づいたベイズスペクトル 密度の方がジェフリーズ事前分布に基づいたベイズスペクトル密度よりも精度よ く推定できることを示している。高次のAR過程での優調和事前分布はTanaka( 2009)によって初めて与えられたが、特性方程式の根を用いた表示のため、数値 実験を行う上では取り扱いづらかった。本発表では高次のAR過程への応用を念頭 において偏自己相関係数(PAC)によるパラメータ表示を導入し数値実験した結 果を紹介する。 また、PACパラメータによる表示は解析的な取扱いをする上でも利点があり、AR 過程の優調和事前分布に関して新しく得られた結果も幾つか紹介したい。
[ Reference URL ]
https://www.ms.u-tokyo.ac.jp/~kengok/statseminar/2009/06.html

### 2009/09/30

15:00-16:10   Room #128 (Graduate School of Math. Sci. Bldg.)

HDLSSデータにおけるPCAについて
[ Abstract ]
マイクロアレイデータなどに見られるように,データの次元数dが標本数nよりも遥かに大きな高次元小標本(HDLSS)データが,解析対象になる場面が増えてきている.

HDLSSデータに対して従来の統計手法を用いると,次元の呪いによって解析が上手くいかない.解決策の一つとして次元縮約法があり, その一つにPCAがある.高次元における従来型のPCAの漸近的性質は,正規性もしくは同等な仮定のもとで,先行研究が多数存在する. しかしながら,これら仮定は,HDLSSを研究する上で,厳しい制約にもなっている. Yata and Aoshimaの一連の研究は,この制約条件の枠を外すことから始まった.HDLSSにおける従来型PCAの限界は何か?推測が一致性をもつための標本数nと 次元数dの関係が,オーダー条件として明らかにされる.従来型PCAの限界を超える手法は何か?一つの実用的な方法として,クロス行列と呼ばれるデータの変換行列が導入され, この行列の特異値分解に基づいた新しいPCAが提案される.

[ Reference URL ]
https://www.ms.u-tokyo.ac.jp/~kengok/statseminar/2009/05.html

### 2009/06/03

15:00-16:10   Room #128 (Graduate School of Math. Sci. Bldg.)

[ Abstract ]

DNA配列多様性と形質多様性との関係を検出するにあたり、このDNA配列集団の多様性と不均一性は、遺伝因子間の非独立性として、関係検出過程に大きな影響を与えることから、DNA配列集団の多様性の把握そのものが、遺伝統計学の課題となっている。
ヒトDNA配列は4種類の塩基が長さ30億であるため、『生命体として成立しうる』という制約の下、非常に多様な配列を取り得る。このDNA配列が取り得る範囲をDNA配列の空間とみなしたとき、DNA配列集団の多様性は、その空間におけるDNA配列集団の分布状態となる。

[ Reference URL ]
https://www.ms.u-tokyo.ac.jp/~kengok/statseminar/2009/04.html

### 2009/05/12

16:20-17:30   Room #126 (Graduate School of Math. Sci. Bldg.)

Asymptitically efficient estimation of multiple change points in GARCH types models
[ Abstract ]
Instability of volatility parameters in GARCH models in an important issue for analyzing financial time series. In this paper we investigate the asymptotic theory for multiple change point estimators of GARCH$(p,q)$ models. When the parameters of a GARCH models have changed within an observed realization, two types estimators, Maximum likelihood estimator (MLE) and Bayesian estimator (BE), are proposed. Then we derive the asymptotic distributions for these estimators. The MLE and BE have different limit laws, and the BE is asymptotically efficient. Monte Carlo studies on the finite sample behaviors are conducted. Applications to Nikkei 225 index are discussed.
[ Reference URL ]
https://www.ms.u-tokyo.ac.jp/~kengok/statseminar/2009/03.html

### 2009/04/22

15:00-16:10   Room #128 (Graduate School of Math. Sci. Bldg.)
Arnaud DOUCET (統計数理研究所)
Interacting Markov chain Monte Carlo Methods for Solving Nonlinear Measure-Valued Equations
[ Abstract ]
We present a new class of interacting Markov chain Monte Carlo algorithms for solving numerically discrete-time measure-valued equations. The associated stochastic processes belong to the class of self-interacting Markov chains. In contrast to traditional Markov chains, their time evolution depend on the occupation measure of their past values. This general methodology allows us to provide a natural way to sample from a sequence of target probability measures of increasing complexity. We develop an original theoretical analysis to analyze the behaviour of these iterative algorithms. We establish a variety of convergence results including exponential estimates and a uniform convergence theorem with respect to the number of target distributions. We also illustrate these algorithms in the context of Feynman-Kac distribution flows.
(this is joint work with Professor Pierre Del Moral)
[ Reference URL ]
https://www.ms.u-tokyo.ac.jp/~kengok/statseminar/2009/02.html

### 2009/04/15

16:20-17:30   Room #128 (Graduate School of Math. Sci. Bldg.)
Jean JACOD (Universite Paris VI)
Estimating the successive Blumenthal-Getoor indices for a discretely observed process
[ Abstract ]
Letting F be a Levy measure whose "tail" $F ([-x, x])$ admits an expansion $\\sigma_{i\\ge 1} a_i/x^\\beta$ as $x \\rightarrow 0$, we call $\\beta_1 > \\beta_2 >...$ the successive Blumenthal-Getoor indices, since $\\beta_1$ is in this case the usual Blumenthal-Getoor index. This notion may be extended to more general semimartingale. We propose here a method to estimate the $\\beta_i$'s and the coefficients $a_i$'s, or rather their extension for semimartingales, when the underlying semimartingale $X$ is observed at discrete times, on fixed time interval. The asymptotic is when the time-lag goes to $0$. It is then possible to construct consistent estimators for $\\beta_i$ and $a_i$ for those $i$'s such that $\\beta_i > \\beta_1 /2$, whereas it is impossible to do so (even when $X$ is a Levy process) for those $i$'s such that $\\beta_i < \\beta_1 /2$. On the other hand, a central limit theorem for $\\beta_1$ is available only when $\\beta_i < \\beta_1 /2$: consequently, when we can actually consistently estimate some $\\beta_i$'s besides $\\beta_1$ , then no central limit theorem can hold, and correlatively the rates of convergence become quite slow (although one know them explicitly): so the results have some theoretical interest in the sense that they set up bounds on what is actually possible to achieve, but the practical applications are probably quite thin.
(joint with Yacine Ait-Sahalia)
[ Reference URL ]
https://www.ms.u-tokyo.ac.jp/~kengok/statseminar/2009/00.html

### 2009/04/15

15:00-16:10   Room #128 (Graduate School of Math. Sci. Bldg.)
Jean JACOD (Universite Paris VI)
A survey on realized p-variations for semimartingales
[ Abstract ]
Let $X$ be a process which is observed at the times $i\\Delta_n$ for $i=0,1,\\ldots,$. If $p>0$ the realized $p$-variation over the time interval $[0, t]$ is

V^n(p)_t=\\sum_{i=1}^{[t/\\Delta_n]}|X_{i\\Delta_n}-X_{(i-1)\\Delta_n}|^p.

The behavior of these $p$-variations when $\\Delta_n ightarrow 0$ (and t is fixed) is now well understood, from the point of view of limits in probability (these are basically old results due to Lepingle) and also for the associated central limit theorem.
The aim of this talk is to review those results, as well as a few extensions (multipower variations, truncated variations). We will put some emphasis on the assumptions on $X$ which are needed, depending on the value of $p$.
[ Reference URL ]
https://www.ms.u-tokyo.ac.jp/~kengok/statseminar/2009/00.html

### 2009/02/04

15:00-16:10   Room #128 (Graduate School of Math. Sci. Bldg.)

Non-Parametric Statistics for a partial sums of iid observations: New Trials
[ Abstract ]
I would lke to review Alpha- quantiles and Ranks with "Empirical Distributions" defined on partial sums of iid. observations discussed in the time continuous version (Brownian Motion).. Then revisiting the formulation of classical estimand and estimators for iid observations , described for example in Fillipova's paper, I would lik,e to discuss on what we could do for our partial sums of iid observations in orde to define our non-parametric estimators and estimands. I will be talking only on the ideas, but mathematical proofs will not be provided.
[ Reference URL ]
https://www.ms.u-tokyo.ac.jp/~kengok/statseminar/2008/13.html

### 2009/02/04

16:20-17:30   Room #128 (Graduate School of Math. Sci. Bldg.)

Black-Scholes 周りの摂動展開について(前半)/ 確率積分の離散化誤差について(後半)
[ Abstract ]
(前半)確率ボラティリティモデルに対して知られている, Black-Scholes モデル周りでの各種摂動展開が統一的にマルチンゲール展開の理論によって 厳密に正当化かつ一般化されることを示す. またとくに拡散過程モデルに 対しては再生法を用いてより精密な結果を与える.

(後半)確率積分の近似として, 増大停止時刻列による区間分割 Riemann 和 をとったとき, その近似誤差の漸近分布を与える. ファイナンスへの応用とし てデルタヘッジエラーを解析し, 取引費用を考慮した上で漸近的に平均2乗誤 差を最小化する戦略を定義する.
[ Reference URL ]
https://www.ms.u-tokyo.ac.jp/~kengok/statseminar/2008/14.html

### 2009/02/04

13:40-14:50   Room #128 (Graduate School of Math. Sci. Bldg.)
Stefano Maria Iacus (Universita degli Studi di Milano)
Applications of Iterated Function Systems to Inference and Simulation
[ Abstract ]
The Iterated Function Systems (IFSs) were born in mid eighties as applications of the theory of discrete dynamical systems and as useful tools for buildings fractals and other similar sets or to produce image compression algorithms. The fundamental result on which the IFS method is based is the Banach contraction theorem because IFSs are defined as operators with some contractive property. In practical applications the crucial point is to solve the inverse problem: given an element f in some metric space (S,d), find a contraction T:S -> S that admits a unique fixed point p such that d(f,p)< eps. When eps=0 the inverse problem is solved exactly and the fixed point p can be identified with the operator T, but in most cases T is an approximation of the target f and T takes linear forms. We present applications of the IFS technique to the problem of estimation of distribution and density functions and to the simulation of L2 stochastic processes.
[ Reference URL ]
https://www.ms.u-tokyo.ac.jp/~kengok/statseminar/2008/12.html

### 2008/12/17

13:40-14:50   Room #002 (Graduate School of Math. Sci. Bldg.)
Ilia Negri (University of Bergamo, Italy)
Goodness of fit tests for ergodic diffusions by discrete sampling schemes
[ Abstract ]
We consider a nonparametric goodness of fit test problem for the drift coefficient of one-dimensional ergodic diffusions, where the diffusion coefficient is a nuisance function which is estimated in some sense. Using a theory for the continuous observation case, we construct two kinds of tests based on different types of discrete observations, namely, the data observed discretely in time or in space. We prove that the limit distribution of our tests is the supremum of the standard Brownian motion, and thus our tests are asymptotically distribution free. We also show that our tests are consistent under any fixed alternatives.

joint with Yoichi Nishiyama (Inst. Statist. Math.)
[ Reference URL ]
https://www.ms.u-tokyo.ac.jp/~kengok/statseminar/2008/09.html

### 2008/12/17

15:00-16:10   Room #002 (Graduate School of Math. Sci. Bldg.)
Stefano Maria Iacus (Universita degli Studi di Milano, Italy)
Divergences Test Statistics for Discretely Observed Diffusion Processes
[ Abstract ]
In this paper we propose the use of $\\phi$-divergences as test statistics to verify simple hypotheses about a one-dimensional parametric diffusion process dXt = b(Xt, theta)dt + sigma(Xt, theta) dWt, from discrete observations at times ti = i*Dn, i=0, 1, ..., n, under the asymptotic scheme Dn - 0, n*Dn - +oo and n*Dn^2 - 0. The class of phi-divergences is wide and includes several special members like Kullback-Leibler, Renyi, power and alpha-divergences. We derive the asymptotic distribution of the test statistics based on phi- divergences. The limiting law takes different forms depending on the regularity of phi. These convergence differ from the classical results for independent and identically distributed random variables. Numerical analysis is used to show the small sample properties of the test statistics in terms of estimated level and power of the test.

joint work with A. De Gregorio
[ Reference URL ]
https://www.ms.u-tokyo.ac.jp/~kengok/statseminar/2008/10.html

### 2008/12/17

16:20-17:30   Room #002 (Graduate School of Math. Sci. Bldg.)
Nicolas Privault (City University of Hong Kong)
Stein estimation of Poisson process intensities
[ Abstract ]
In this talk we will construct superefficient estimators of Stein type for the intensity parameter lambda > 0 of a Poisson process, using integration by parts and superharmonic functionals on the Poisson space.
[ Reference URL ]
https://www.ms.u-tokyo.ac.jp/~kengok/statseminar/2008/11.html