next up previous contents
Next: B. Algorithms Up: A. Statistical tools Previous: A.2 Maximum likelihood


A.3 Iterative mean

The mean value of a distribution {xi} can also be computed iteratively if the values xi are drawn one-by-one. Let $ \left\langle\vphantom{ x }\right.$x$ \left.\vphantom{ x }\right\rangle_{t}^{}$ be the average over the first t data points. We observe that


$\displaystyle \left\langle\vphantom{ x }\right.$x$\displaystyle \left.\vphantom{ x }\right\rangle_{t}^{}$ = $\displaystyle {\frac{{1}}{{t}}}$$\displaystyle \sum_{{i=1}}^{t}$xi  
  = $\displaystyle {\frac{{1}}{{t}}}$$\displaystyle \sum_{{i=1}}^{{t-1}}$xi + $\displaystyle {\frac{{1}}{{t}}}$ xt  
  = $\displaystyle {\frac{{t-1}}{{t}}}$ $\displaystyle {\frac{{1}}{{t-1}}}$$\displaystyle \sum_{{i=1}}^{{t-1}}$xi + $\displaystyle {\frac{{1}}{{t}}}$ xt  
  = $\displaystyle {\frac{{t-1}}{{t}}}$$\displaystyle \left\langle\vphantom{ x }\right.$x$\displaystyle \left.\vphantom{ x }\right\rangle_{{t-1}}^{}$ + $\displaystyle {\frac{{1}}{{t}}}$ xt . (A.13)

Thus, the update rule for the temporary mean c(t + 1) = $ \left\langle\vphantom{ x }\right.$x$ \left.\vphantom{ x }\right\rangle_{t}^{}$ upon presentation of a value x is

c(t + 1) = c(t) + $\displaystyle {\frac{{1}}{{t+1}}}$$\displaystyle \left(\vphantom{x - c(t)}\right.$x - c(t)$\displaystyle \left.\vphantom{x - c(t)}\right)$ . (A.14)

This rule is the same as for the center update in vector quantization with a learning rate that decays as 1/t.


next up previous contents
Next: B. Algorithms Up: A. Statistical tools Previous: A.2 Maximum likelihood
Heiko Hoffmann
2005-03-22