I'm not aware of any papers explicitly testing this, but here are some thoughts:

• IgG: I agree with you on this, it's essentially the noise in the experiment. Trying to look at enriched regions by doing a log fold change would result in dividing by close-to-zero, which doesn't make sense. Subtracting the IgG, rather than dividing, would make more sense, but you still wouldn't have a good estimate of the chromatin going into the IP.

• Untagged strain: this could be good, but I'd argue that this is introducing another variable (i.e., the culturing of the additional strain) that's not controlled for.

• Input: I think this is the best option. It gives you exactly the chromatin that was available for the IP. Furthermore, peak-calling algorithms are typically written with assumptions appropriate for input as the control. Any possible biological changes to the chromatin induced by the HA-tagged protein will show up in the input, something that you won't see in the untagged strain.

I suppose if you have tons of money to spend and tons of starting material, you could do (IP - IgG) to hopefully control for non-specific binding, and then compare that to input as control -- though I'm not aware of anyone who's actually done this.

I found that this recent review gave a good overview for chip-seq including controls. http://www.nature.com/ni/journal/v12/n10/abs/ni.2117.html

I think most labs use Input because IgG can be biased (I'm working w/human cells and I have not heard of untagged strain). Excerpt from the above paper:

"IgG may be less desirable in certain circumstances because of the following reasons: most IgG antibodies are not obtained from true preimmune serum from the same animal in which the specific antibody was raised; and IgG antibodies usually immunoprecipitate much less DNA than specific antibodies do, and thus limited genomic regions from the control may be overamplified during the library construction step"