{%hackmd aPqG0f7uS3CSdeXvHSYQKQ %} # Power Method ### Example 看一下費波那契數列 $F_n = F_{n-1} + F_{n-2}$,這個數列長 $\{0,1,1,2,3,5,8,13,21,34,\ ...\}$,我們可以透過矩陣來重寫這個數列: <center> <img src = "https://github.com/Mes0903/MesBlog/blob/vuepress-theme-hope/src/numerical_algebra/Power-Method/image/1.png?raw=true"> </center><br> 在這個例子裏面,我們觀察到 eigenvector、eigenvalue 跟長期的外顯行為有關。 ### Example 看第二個例子,假設這邊有四個網頁,然後我們常像這樣去瀏覽她: <center> <img src = "https://github.com/Mes0903/MesBlog/blob/vuepress-theme-hope/src/numerical_algebra/Power-Method/image/2.png?raw=true"> </center><br> 那麼就會有一個機率矩陣: <center> <img src = "https://github.com/Mes0903/MesBlog/blob/vuepress-theme-hope/src/numerical_algebra/Power-Method/image/3.png?raw=true"> </center><br> ### Thm dominant eigenvalue <center> <img src = "https://github.com/Mes0903/MesBlog/blob/vuepress-theme-hope/src/numerical_algebra/Power-Method/image/4.png?raw=true"> </center><br> ### Prop Power Method 假設 $A \in \mathbb{R}^{n \times n}$ 有 dominant eigenvalue。 給定 initial guess $\vec x$ 且製造一個數列長這樣: <center> <img src = "https://github.com/Mes0903/MesBlog/blob/vuepress-theme-hope/src/numerical_algebra/Power-Method/image/5.png?raw=true"> </center><br> 我們希望這個數列能很好的幫助我們去逼近 dominant eigenvalue。 ### Example <center> <img src = "https://github.com/Mes0903/MesBlog/blob/vuepress-theme-hope/src/numerical_algebra/Power-Method/image/6.png?raw=true"> </center><br> 這裡我們可以看到 Power Method 會產生一個很大的數字在矩陣前方,這個數字可以透過歸一化之類的方法來消除掉,我們這邊透過 scale down 的方法來做一次,在每步迭代前都先除上自己 norm: <center> <img src = "https://github.com/Mes0903/MesBlog/blob/vuepress-theme-hope/src/numerical_algebra/Power-Method/image/7.png?raw=true"> </center><br> ### Thm dominant eigenvalue <center> <img src = "https://github.com/Mes0903/MesBlog/blob/vuepress-theme-hope/src/numerical_algebra/Power-Method/image/8.png?raw=true"> </center><br> 證明: <center> <img src = "https://github.com/Mes0903/MesBlog/blob/vuepress-theme-hope/src/numerical_algebra/Power-Method/image/9.png?raw=true"> </center>
×
Sign in
Email
Password
Forgot password
or
By clicking below, you agree to our
terms of service
.
Sign in via Facebook
Sign in via Twitter
Sign in via GitHub
Sign in via Dropbox
Sign in with Wallet
Wallet (
)
Connect another wallet
New to HackMD?
Sign up