I find myself doing searches like this a lot. My method is basically depth first search, but I don't claim this is optimal. Anyway, I know David knows what DFS is, but I'll describe my algorithm is detail anyway: First, I'll try to remember where I've read something like this before or start googling till I find something in the ballpark of the result I want. If that paper gives references to other sources for background material (e.g. textbooks, etc) then I'll go look them up. I keep tracing back the chain of references till I am pretty sure the result is not there. Then I go up the chain of papers till I find another avenue which seems fruitful and check it out. Because I keep digital copies of almost all papers and books it's easy to search within them for the keywords. The benefit of doing it this way is that if the terminology has changed at some point in history, and if the papers I'm tracing back along are well-written, then I'll be made aware of which keywords to search for in older documents.

I often supplement this process above by adding keywords or phrases to my original google search to pare down the number of hits. I never thought of it like this before, but this is the triangulation part of the search. The idea in my mind doing this is that if a paper is going to contain the result I want then it will also probably contain this phrase. If I know it should contain the definition of a foobar then I might add the phrase "a foobar is" or "is called a foobar" to the googling. Sometimes I'll have multiple tabs open with different google search results. Often I can get one of them down to under 30 hits. Because there's some choice in this process I often don't get the original place the result appeared, but I can find a place where it's there as a lemma or something. And if I really care about the original place then I can follow the references in that paper and usually find it, or copy the verbiage of that lemma to see where the author got it from.

I find myself doing searches like this a lot. In the pre-google days I imagine my method would be basically depth first search through the references, but I don't claim this is optimal. Anyway, I know David knows what DFS is, but I'll describe what I mean in detail anyway because it's the basis for what I really do. First, I'll try to remember where I've read something like this before or start googling till I find something in the ballpark of the result I want. If that paper gives references to other sources for background material (e.g. textbooks, etc) then I'll go look them up. I keep tracing back the chain of references till I am pretty sure the result is not there. Then I go up the chain of papers till I find another avenue which seems fruitful and check it out. Because I keep digital copies of almost all papers and books it's easy to search within them for the keywords. The benefit of doing it this way is that if the terminology has changed at some point in history, and if the papers I'm tracing back along are well-written, then I'll be made aware of which keywords to search for in older documents.

Since we live in a world with google, I often supplement this process above by adding keywords or phrases to my original google search to pare down the number of hits. I never thought of it like this before, but this is the triangulation part of the search. The idea is that if a paper is going to contain the result I want then it will also probably contain this phrase. If I know it should contain the definition of a foobar then I might add the phrase "a foobar is" or "is called a foobar" to the googling. I usually add enough phrases, one step at a time, till the number of results is down to between 5 and 30 hits.

Obviously, there are lots of choices being made in which phrases to add, so I have to estimate probabilities on the fly to guess at which choice is going to be most likely to pare down the results without throwing away the paper that contains the reference I want. I guess this would be the hardest part for a machine to emulate. I make my estimates based on all the papers I've read and the way in which authors usually write. Computers can do things like this nowadays, e.g. the algorithms used to summarize the news. I keep multiple tabs open with different google search results so I can try several of the most likely searches. The tabs usually have some overlap but also some unique results. In the end I don't always get the original place the result appeared, but I can find a place where it's there as a lemma or something. And if I really care about the original place then I can follow the references in that paper and usually find it, or copy the verbiage of that lemma into google to see where the author got it from.