There's some pretty good evidence that numerical cognition emerged fairly early in the primate lineage, at least, if not significantly earlier in evolution. Most of the work on numerical cognition in non-human primates, however, has focused on a handful of monkey and ape species. The prosimian suborder of primates, however, which includes lemurs, diverged from the main primate lineage some 47-54 million years ago. If the numerical abilities of lemurs parallel the numerical abilities of other non-human primates (such as rhesus macaques or chimpanzees, for example), then it may be possible to infer that numerical cognition in emerged prior to the divergence of lemurs, lorises, and galagos (the prosimians). Until just a few years ago, there had been no systematic research of numerical cognition in lemurs.
Well, Elizabeth Brannon of Duke University, and her colleagues, decided to put an end to that. They set up an investigation of numerical cognition in mongoose lemurs (Eulemur mongoz) at the Duke Lemur Center.
In order to address this question, the nine lemur participants watched as the experimenters placed grapes into a green bucket on the floor of their enclosures. On some of the trials, the lemur was allowed to retrieve all of the grapes from inside the bucket, after the experimenter had finished. On other trials, however, the experimenters hid several of the grapes into a hidden compartment in the bottom of the bucket, making them irretrievable. If the lemurs were able to count the number of grapes that they observed being placed into the bucket, then they should search longer during the trials when they encountered the apparent "missing" grapes.
The bars in this graph reflect the difference in time spent comparing the full-retrieval trials and the partial-retrieval trials. For example, the first bar reflects extra time searching for the second grape in a 2-grape trial when only one grape was retrievable (and the second grape was hidden in the secret compartment).
When the ratio of grapes retrievable to expected differed along a 1:2 ratio, the lemurs spent extra time searching. That suggests that mongoose lemurs are able to numerically distinguish sets that differ by a factor of two. However, they did not spend extra time looking when the ratio of grapes retrievable to expected differed along a 2:3 or 3:4 ratio. In other words, even if they had seen 4 grapes disappear into the bucket, for example, upon only finding three of them, they were satisfied and discontinued their search. From this, we can infer that mongoose lemurs are unable to numerically distinguish sets of objects that differ by a factor of 2:3 or 3:4
While this data strongly suggests that lemurs are able to distinguish sets at a 1:2, but not 2:3 or 3:4 ratio, there are two confounding variables that must be considered.
It is possible that lemurs rely on olfaction more than vision, which leads to longer search times when grapes are hidden in the secret compartment. While this is unlikely (given the results of the 2:3 and 3:4 conditions), an additional control condition was included in the experiment. In this case, the lemurs watched as one grape was hidden in the bucket, and they were allowed to retrieve the same one grape. However, six grapes were previously hidden in the secret compartment, outside the sight of the lemurs. If they were using olfaction rather than vision, they should have continued to search for the hidden grapes. In fact, they did not, reflecting the notion that the lemurs were using vision rather than olfaction to count and retrieve the grapes.
It is also possible that the lemurs were using some variable that correlates with number, but not number itself, to track the amount of grapes in the bucket. For example, it is possible that they could estimate the volume of grapes hidden, rather than the number of grapes hidden, and this could lead to errors in grape retrieval. To address this possibility, the lemurs observed the experimenters place two grape halves into the bucket (they were actually hidden in the secret compartment), and then the lemur was able to retrieve a single, whole grape that had already been in the bucket previously. If the lemurs were tracking the volume of grapes hidden, then they should have been satisfied upon retrieving the whole grape and discontinued their search. Instead, the lemurs continued their search, inferring that the whole grape had been in the bucket the whole time, and thinking that the two grape halves must have been in their somewhere.
Taken together, these data strongly suggest that early prosimian primates, such as mongoose lemurs, have mental representations of number, and that they are able to distinguish among sets that differ on a 1:2 ratio, but not a 2:3 or 3:4 ratio. The fact that these early primates possess these numerical abilities may not be entirely surprising, as similar abilities have been seen in a wide array of animals, including chickens and fish.
Indeed, numerical cognition is one of the most well-fleshed out lines of research in comparative cognition today. The main challenge moving forward, as the authors indicate, will be to attempt to determine whether the core knowledge systems for number emerged extremely early and are the same across multiple species and phyla, or whether they emerged multiple times in different lineages in a process of convergent evolution, due to similar ecological or social pressures.
Lewis KP, Jaffe S, & Brannon EM (2005). Analog number representations in mongoose lemurs (Eulemur mongoz): evidence from a search task. Animal cognition, 8 (4), 247-52 PMID: 15660208
Lemur image source.
If the animals were food deprived prior to being placed in the experimental condition, the task might become more salient and the results clearer.
"If the numerical abilities of lemurs parallel the numerical abilities of other non-human primates (such as rhesus macaques or chimpanzees, for example), then it may be possible to infer that numerical cognition in emerged prior to the divergence of lemurs, lorises, and galagos (the prosimians)."
How does this work? Humans and New Caledonian Crows both make tools, but humans are synapsids and crows are diapsids. Are you suggesting that tool making can therefore be inferred in stem amniotes?
I'm not an expert on numerical cognition or evolution, but I think the difference is that in the human/crow case, there would need to have occurred many losses of numerical cognition (in other words, every branch of the amniote cladogram that did NOT lead to either humans or crows would need to have lost the trait). Each of those trait losses makes that explanation less parsimonious, which is why we consider it much more likely that this is a case of convergent evolution (numerical cognition evolved twice, in corvids and in primates). In the human/lemur case, the idea is that N.C. evolved once and is present in all or most of the species descended from that ancestor. Can you think of a reason this doesn't work?