スキップしてメイン コンテンツに移動

ヘッセ行列

Introduction


今日は、ヘッセ行列を用いたテイラー展開について書こうと思います。
これは最適化を勉強するにあたって、とても大事になってくるので自分でまとめて残しておくことにしました。とくに、機械学習では最適化を必ず行うため、このブログのタイトルにもマッチした内容だと思います。
.

概要

  • ヘッセ行列の定義
  • ベクトルを用いたテイラー展開
  • 関数の最適性

ヘッセ行列の定義

仮定

f は次のような条件を満たす関数です。.
  • f はn次元ベクトルから実数値を出力します。
    このベクトルは次のように表せます。
    \[x = [x_1,x_2,,,,x_n]\]
  • \(\forall x_i , i \in {1,2,,,n}\), f は二回偏微分可能です。

定義

ヘッセ行列は \(\frac{\partial^2}{\partial x_i \partial x_j}を (i,j)要素に持ちます。\)
よってヘッセ行列は次のように表せます。
\[ H(f) = \left( \begin{array}{cccc} \frac{\partial^ 2}{\partial x_1^2} & \frac{\partial^2 f}{\partial x_1 \partial x_2} & &\ldots \frac{\partial^2 f}{\partial x_1 \partial x_n} \\ \frac{\partial^ 2 f}{\partial x_1 \partial x_2} & \frac{\partial^ 2 f}{\partial x_2^ 2} & \ldots & \frac{\partial^2 f}{\partial x_2 \partial x_n} \\ \vdots & \vdots & \ddots & \vdots \\ \frac{\partial^ 2 f}{\partial x_n \partial x_2} & \frac{\partial^ 2 f}{\partial x_n \partial x_2} & \ldots & \frac{\partial^2 f}{\partial x_n^2} \\ \end{array} \right) \]

ベクトルを用いたテイラー展開

テイラー展開を二次の項まで行った式を与えます。
\[f(a+h) = f(a) + \nabla f(a) h + \frac{1}{2} h^T \nabla f(a) h + R_3\]
ただし、 \(H=\nabla ^2 f\) はヘッセ行列です。
大事なのは二次の項までで止めていることです。理由は最適化できるようにです。

関数の最適性

定置性

  • \(n \times n\)行列A が 正定置
    \(\forall x \in\)n次元ベクトル空間、\(z^T A z > 0\)
  • \(n \times n\)行列A が 負定置
    \(\forall x \in\)n次元ベクトル空間、\(z^T A z < 0\)
  • \(n \times n\)行列A が 半正定置
    \(\forall x \in\)n次元ベクトル空間、\(z^T A z => 0\)
  • \(n \times n\)行列A が 半不定値
    \(\forall x \in\)n次元ベクトル空間、\(z^T A z <= 0\)
\(z^T A z\) は二次形式と呼ばれています。

最適性

もちろん上記で書いた二次の項までのテイラー展開の式は二次形式を含みます。
よって関数の最適性は次のように考えられます。
  • H(a)(ヘッセ行列) が正定置\(\implies\)f(a) は極小値.
  • H(a)(ヘッセ行列) が負定置\(\implies\) f(a) は極大値

Reference

https://ja.wikipedia.org/wiki/%E3%83%98%E3%83%83%E3%82%BB%E8%A1%8C%E5%88%97
http://www2.kaiyodai.ac.jp/~takenawa/optimization/resume10-1.pdf
http://www.dais.is.tohoku.ac.jp/~shioura/teaching/mp04/mp04-8.pdf
http://tau.doshisha.ac.jp/lectures/2008.calculus-II/html.dir/node43.html

コメント

このブログの人気の投稿

Thorem of SVM part 2

Introduction 日本語 ver This post is written about theorem of SVM. This post is part 2. Part 1 is written about deriving the objective function. I will write about dual problem. The objective function which deriving in Part 1 is called the main problem. Typically, We optimize not the main problem, but the dual problem. I will write about deriving the dual problem from the main problem. If you until look up part 1, please look up Theorem of SVM part 1 . I implement SVM. It post is Implement linear SVM Implement kernel SVM Overview Main problem Dual problem  Lagurange function maximize L about dual variable, minimize L about primal variable minimize L about primal variable, maximize L about dual variable. Main problem I will review the main problem in Part 1. $$\min_{w,b} \frac{1}{2}||W||^2 + C\sum_{i \in N} \epsilon_i$$ $$~~s.t~~ \forall i \in N, y_i (w^T \phi(x_i) + b) \geq 1 - \epsilon_i ,~~~, \forall i \in N~\epsilon \geq 0$$ This problem is called s...

K-means 理論編

Introduction English ver 今日はK-meansアルゴリズムの理論について書きます。 K-meansアルゴリズムはクラスタリングのためのアルゴリズムです。 K-meansの実装の記事は カーネルK-meansの実装 を御覧ください。 この記事はカーネルK-menasの実装についての記事ですが、通常のK-meansの実装も行っています。カーネルK-meansについてはまた、今度別の記事で紹介したいと思います。 概要 1 of K 符号化法 プロトタイプ 歪み尺度 最適化 1 of K 符号化法 K-meansはK個のクラスについて分類することを考えます。 K-meansでは $x_n$がkのクラスに属していることを次のように表します。 ベクトル$r_n:1 \times K$ を $$r_n := (0,0,..,1,..,0)$$ このベクトルはk番目にのみ1を持ち、それ以外は0を要素に持つようなベクトルです。 こののような表現の仕方を1 of K符号化法と呼びます。 プロトタイプ K-meansではプロトタイプと呼ばれるベクトルを選びます。このベクトルは各クラスに一つあり、そのクラスの代表のようなベクトルです。 K-means ではそのようなベクトルは各クラスの平均ベクトルとなります。これは目的関数から自然と導かれます。 歪み尺度 プロトタイプベクトルを $\mu_i ~\forall k \in K$とします。 この時、k-meansの目的関数は次のようになります。 $$J = \sum_{n=1}^{N} \sum_{k=1}^{K} r_{nk} ||x_n-\mu_k||^2$$ ここで、 $r_{nk}$ は$r_n$のk番目の要素です。 この目的関数について少し説明をします。$r_{n}$は$x_n$が属しているクラスのラベルの場所だけ1で他は0であるので、 $$J = \sum_{n=1}^{N} ||x_n - \mu_{x_n}||$$ ここで、$\mu_{k_n}$は$x_n$が属しているクラスのプロトタイプです。 よって、 $$J = ||x_1 - \mu_{x_1}|| + ||x_2 -\mu_{x_2}|| + ...

Theorem of K-means

Introduction 日本語 ver Today, I will write about the theorem of K-means algorithm. K-means algorithm is a method to do clustering about K of class. The post of implement K-means is Implement kernel K-means This post is written about kernel K-means. The kernel K-means is the method which covers the weak point of K-means. I will explain kernel K-means another post. Overview 1 of K coding scheme prototype vector Distortion measure Computing 1 of K coding scheme K-means algorithm is clustering K of class. Now, K-means algorithm expresses that $x_n$ be belong to k like as follows. Let vector $r_n:1 \times K$ is $$r_n := (0,0,..,1,..,0)$$ this vector have 1 in element of k'th and have 0 in else. This expression is called 1 of K coding scheme. Prototype vector K-means algorithm chooses vector which called a prototype. this vector has represented by the cluster. K-means algorithm is regard mean vector as representative of the cluster. I...