You have limited data and you must use some statistical methods to allow for confidence that the expression effects seen are not random variations and actually associated with the biology of the samples investigated. Usually we recommend to run each condition in duplicates or two technical replicates, so that any changes that are observed are actually significant. So on a dual colour system, ratio of control versus test on two different arrays is generally accepted as statistically significant.
To answer your second part:
- Replicates can be used to measure variation and permit statistical tests to evaluate the variation.
- Averaging of the replicates can increase precision and enhanced detection of differences in gene expression
- Enable detection of outliers in a set
The comparative analysis first requires some intensity normalization of the samples so that there is a common starting point. Next the use of statistical methods such as t-statistics and ANOVA may be employed to determine variance and assign a p-value (essentially a value that suggests the confidence in the measurement) to each comparative result. The p-value threshold may be adjusted in order to obtain a reasonable number of top hits (e.g., greater than 100). You are essentially trying to overcome two things 1) distinguish what is "real" from what is random, and 2) avoid finding false patterns in data. The p value helps make this effort easier.
Microarray results are not directly quantitative. So if one finds a genes that is some number >1.5 or 2 fold or more elevated or depressed in their comparative expression results further work follows. Most microarray experimental results are usually confirmed by other methods such as by some form of quantitative PCR on the genes though to be most associated with the results.