Ultimately, there has to be functional data showing that a particular mutation can transform cells. But realistically, we will never be able to test all of the potential drivers in all types of tumor in that way.

From a genomics perspective, the key question is whether a particular basepair/gene/locus undergoes somatic changs at a greater rate than chance would predict. If it does, then we can suppose that change confers a selective advantage and is likely a driver.

Current "state of the art" models like MutSig for coding changes:

http://www.broadinstitute.org/cancer/cga/mutsig

and GISTIC for copy number changes:

http://www.broadinstitute.org/software/cprg/?q=node/31

attempt to do just this. These approaches are not perfect but I think they've helped us take a step forward in separating drivers from passengers.

Recurrence doesn't necessarily mean its a driver mutation. You would be more confident to call something a driver mutation if you were able to use it, for example, to transform cells or create tumours using transgenic models.

A driver mutation gives a selective advantage to a clone through either increasing its survival or reproduction. Driver mutations usually tend to cause clonal expansions.

I am not sure there is a clear answer to that. One tool that can be easily used is to check if a mutation is listed in the catalogue of cancer gene mutations at COSMIC. In the absence of any prior information, that one would have to come up with an experimental approach, which as mentioned would have to vary depending on where the mutation is found. 

Ultimately, there has to be functional data showing that a particular mutation can transform cells. But realistically, we will never be able to test all of the potential drivers in all types of tumor in that way.

From a genomics perspective, the key question is whether a particular basepair/gene/locus undergoes somatic changs at a greater rate than chance would predict. If it does, then we can suppose that change confers a selective advantage and is likely a driver.

Current "state of the art" models like MutSig for coding changes:

http://www.broadinstitute.org/cancer/cga/mutsig

and GISTIC for copy number changes:

http://www.broadinstitute.org/software/cprg/?q=node/31

attempt to do just this. These approaches are not perfect but I think they've helped us take a step forward in separating drivers from passengers.

Pawel give a good answer. Recurrence mutation does not mean that is a driver mutation, however, it may also indicate that specific tumor share the similar molecular pathway.

Amadeu also give a good comment that "A driver mutation gives a selective advantage to a clone through either increasing its survival or reproduction." This may be true in first glance. Bur think about most of the mutations that mute the tumor suppressor gene or activate oncogenes share the similar behavior that give the tumor growth advantage and evade apoptosis. Thus that may not be sufficient to define all of them as driver mutations.

Any one can give a clear and accurate definition for driver mutation?

According to the two step theory, if only the particular one that pass the two steps could be defined as  driver mutation? How about others happened following the first one?

You have received some good answers above.  There isn't ONE good way to determine whether a given genetic event (mutation, deletion, amplification, etc.) is a driver or not. It requires a series of experiments and some bioinformatics.  For example, you could try to see if the mutant allele confers tumorigenic properties on a normal cell of origin (HMEC for breast, ovarian or tubal cells for ovarian cancer, etc..).  Positive results would not necessarily make it a driver or an oncogene.  You would also want to show that silencing the allele in a tumor cell line that harbors it leads to some attenuation of tumor behavior.  You might also want to see (if possible) how early that mutation occurs in the progression from normal epithelium to malignant epithelium.  For instance, in colon cancer you could ask whether the mutation occurs in dysplastic polyps before the invasive carcinoma forms or later.  In high-grade serous ovarian cancer, you could ask whether the mutation occurs in the newly described precursors in the fallopian tube (i.e. p53 signatures and/or serous tubal intraepithelial carcinomas (STICs)) or only in the invasive carcinoma.  It could be argued that early events are more likely to be drivers than later events. 

Driver mutation can be identify on the basis of frequency of that mutant gene, in the large number of samples and its functional relevance.