It says that each cell must be at least 5.

No. The EXPECTED number of observations must be >= 5. It depends on the null

hypothesis, and your assumptions, and/or the study design (e.g. experimental study

with groups of equal size).

For example, if one has 2 x 4 table, including a binary outcome with expected probabilites

0.5 / 0.5, and four experimental groups with proportions 0.25 / 0.25 / 0.25 / 0.25, and a

null hypothesis of independence between group and outcome, then the expected proportion

in each cell would be 0.25 * 0.5 = 0.125. Now it is easy to calculate the total number

of observations required in order to achieve >=5 expected frequencies in each cell.

With kind regards

Karabiner