我认为LOESS中多项式回归其实不重要，重要的是"局部"。在行为科学的统计中，True Model通常不是 Exact Model，而是近似模型。
LOESS因为“局部”的特点，可以收敛到Population而不是近似模型空间。这样就可以通过足够的数据量来衡量模型空间和Population之间的切合程度。King, Tomz, & Wittenberg (2000) 用predicted指模型拟合值，用expected指无模型(Local)拟合值--
Depending on the issue being studied, the expected or mean value of the dependent variable may be more interesting than a predicted value. The difference is subtle but
important. A predicted value contains both fundamental and estimation uncertainty, whereas an expected value averages over the fundamental variability arising from
sheer randomness in the world, leaving only the estimation uncertainty caused by not having an infinite number of observations. Thus, predicted values have a larger variance than expected values, even though the average should be nearly the same in both cases.
参考：
?TeachingDemos::loess.demo
King, Tomz, & Wittenberg (2000), American Journal of Political Science, 44, 347-361.
http://links.jstor.org/sici?sici=0092-5853%28200004%2944%3A2%3C347%3AMTMOSA%3E2.0.CO%3B2-S
