lbwei space

Linear LMS

#Probability
2022/11/29

Table of Content


Motivation

因為 LMS 可能不好解,所以用 LLMS 找近似最佳解。


Solve the LLMS problem

\(\Theta\) 是 prior、\(X\) 是 observation。

我們的 estimator 是線性的,即 \(\hat \Theta = aX+b\)。所以要最小化的目標是 \(E\big[(\Theta-aX-b)^2 \big]\)。

假設 \(a\) 已知,我們可以將 \(\Theta-aX\) 視為一全新的 r.v. \(Y\);使 \(E\big[(Y-b)^2 \big]\) 最小化的 \(b\) 正是 \(E[Y]\) 。於是 \(b = E[\Theta - aX]\)。

現在只剩下一個變數要處理了!我們要最小化的目標變成

\[\begin{align*} E\big[(\Theta-aX-E[\Theta-aX])^2 \big] &= var(\Theta-aX) \\ &= var(\Theta) + a^2var(X) - 2a\cdot cov(\Theta, X)。 \end{align*}\]

這是 \(a\) 的二次函數;因為 \(var(X)\) 恆正(設非零),所以對此微分可以得到最小值所在,也就是

\[a = {cov(\Theta, X)\over var(X)}。\]

總結以上,

\[\begin{align*} \hat \Theta_L &= aX+b \\ &= E[\Theta] + {cov(\Theta, X)\over var(X)}(X-E[X]) \\ &= E[\Theta] + \rho{\sigma_{\Theta}\over \sigma_X}(X-E[X])。 \end{align*}\]

Remark

相較於 LMS 的 \(\hat \Theta = E[\Theta\vert X]\),\(\hat \Theta_L\) 只和 means, variances, and covariance 有關,代表我們不需要知道全部有關 \(\Theta\) 和 \(X\) 的 distribution 資訊。這是很好的性質!

也就是說,對於兩組隨機變數 \(\Theta_1, X_1\) 和 \(\Theta_2, X_2\),只要兩兩對應的 mean, variance, and covariance 都相等,則 \(\hat \Theta_{L, 1} = \hat \Theta_{L, 2}\)!\(\Theta_1, X_1, \Theta_2, X_2\) 是什麼 distribution 不重要!

關於 performance,

\[E\big[(\hat\Theta_L - \Theta)^2 \big] = (1-\rho^2)var(\Theta)。\]

可以先假設 \(E[\Theta]=E[X]=0\),再將等式左邊依定義展開,即可得到等式右邊。由此可見,\(\rho\) 在 LLMS 中扮演重要的角色:不僅影響 estimator,也牽動著誤差。

Exercise

MIT OCW.


Different Data Represnetations

In LMS, \(E[\Theta\vert X]\) is the same as \(E[\Theta\vert X^3]\).

I suspect that once \(g(X)\) is one-to-one and onto, \(E[\Theta\vert X] = E[\Theta\vert g(X)]\), but it is not proved yet.

However, in LLMS, estimator \(\hat \Theta=aX+b\) is different from estimator \(\hat \Theta=aX^3+b\). Which is the better one? That depends on the problem we are solving.

Moreover, we can even consider \(\hat \Theta=a_1X+a_2X^2+a_3X^3+b\). Although the functions of \(X\) appeared are not necessarily linear, the coefficients are linear, and that’s pretty enough. After all, when dealing with LLMS, we solve a system of linear equations of the coefficients, not \(X\)s. To generalize, we can apply the methodology of LLMS for estimators of the form:

\[\hat \Theta_L = b + \sum_{i=1}^na_ig_i(X),\]

where \(g_i(X)\) can be any well-defined function.

越複雜的 estimator 越難計算,但或許效果比較好,視問題而定。


Reference