I’m not sure there are any specific criteria as such – it’s more about typological likelihood and possible ways to explain all the available material: are the attested reflexes likely to develop from each other, or is it more likely that they all developed from a third element? Is it even possible to map out a way in which the attested reflexes developed from each other, or do we run into insurmountable issues if we try?
The best way to illustrate this is probably by way of examples.
One reflex is original
For an example of the former, let’s take the Indo-European short vowels e, a and o in a few languages (speaking in very broad and simplified terms for illustration here, and ignoring lots and lots of actual complications, not least laryngeals):
- In Latin, Greek and Celtic, we see reflexes with /e a o/
- In Germanic, we see reflexes with /e a/, but no /o/
- In Sanskrit and Avestan (Indo-Iranian), we see reflexes with /a/, but no /e o/
So how do we explain this?
Do the Indo-Iranian languages reflect the original state (only one vowel), and that one vowel split into multiple vowels in the others through conditioning? No, that doesn’t seem likely, because we have examples of different vowels in the same environments, and there doesn’t seem to be any possible conditioning that could satisfactorily explain such a split. The same is true if we take the Germanic state to be original: no conditioning adequately explains the split of /a/ into /a/ and /o/ in Latin, Greek and Celtic.
Perhaps Latin, Greek and Celtic reflect the original state (three vowels), and some or all of them merged into one in the others. This seems more likely. We know from attested language changes that a merger of /a/ and /o/ is not uncommon, so Germanic doesn’t pose a problem. We also know from contemporary descriptions of Sanskrit that what we transcribe as /a/ was really more like /ɐ/ or /ə/ – a centralised, somewhat neutral vowel; and we know, again from attested changes, that neutralisation of multiple vowels into a centralised vowel is also quite common. So Indo-Iranian is also explainable.
That means we can actually quite satisfactorily explain this vowel inventory in terms of an attested reflex being (more or less) unchanged from the proto-language.
No reflex is original
As an example of a case where such an explanation doesn’t work, let’s take the Indo-European unvoiced velars (as we now call them). Here we’ll start just with some reflexes that clearly correspond to one another across branches:
- Words where, on the one hand, Greek, Latin and Celtic have /k/ and Germanic has /χ/; on the other, Indo-Iranian has /k/ or /c/ and Balto-Slavic has /k/ or /tʃ/, depending on the following reconstructed vowel, as per above
- Words where Greek has /p/ or /t/ depending on the following reconstructed vowel, Latin and Celtic have /kʷ/, and Germanic has /χʷ/; while Indo-Iranian and Balto-Slavic have the same reflexes as in 1
- Words where Greek, Latin, Celtic and Germanic have the same reflexes as in 1; while Indo-Iranian and Balto-Slavic have various sibilants (roughly Sanskrit /ɕ/, Avestan /s/, Balto-Slavic /ʃ/)
There is one part of this that we can dispense with, and that’s the Germanic /χ χʷ/: we know from lots of other evidence that Germanic underwent a series of changes known as Grimm’s Law, which turned all unvoiced stops into fricatives. So we can treat Germanic /χ χʷ/ as reflecting /k kʷ/, exactly in line with Latin and Celtic.
#1 above also seems fairly straightforward: we essentially have /k/ everywhere, except in Indo-Iranian and Balto-Slavic in cases where originally a front vowel /e i/ followed, in which case we get some sort of palatal sound like /c/ or /tʃ/. Palatalisation of velars before front vowels is one of the commonest sound changes found across the world’s languages, so we can be fairly confident that the ‘western’ group – Greek, Latin, Celtic and Germanic (with Grimm’s Law) – have retained the original unchanged, while the ‘eastern’ group – Indo-Iranian and Balto-Slavic – palatalised the consonant before any subsequent vowel mergers. The proto-sound was /k/.
#2 and #3 are more complicated.
In #2, we have similar evidence of palatalisation in the eastern group, so let’s assume /k/ is the base sound there as well. The western group shows /kʷ/, except for Greek, which has /t/ before front vowels and /p/ elsewhere (but also /k/ in some specific contexts, like between rounded/back vowels). Latin, Celtic and Germanic all agree on /kʷ/, and a change from /kʷ/ to /p/ is quite common, so Greek /p/ is not particularly surprising; the same is true of /kʷ/ to /k/ next to other labial sounds. A change from /kʷ/ to /t/ due to palatalisation is not common, but the evidence overwhelmingly suggests that that is what happened in Greek.1 So #2 is basically western /kʷ/ vs eastern /k/.
In #3, we have almost the reverse scenario: the western group shows the same reflexes as in #1, so we’ll assume western /k/. But the eastern group shows something that’s neither velar nor a stop – a variety of sibilants! The sibilants vary (some are palatal, some are not), and we know that palatalisation and depalatalisation of sibilants are both fairly common changes – so for the time being, let’s assume the eastern group just represents /s/, i.e., #3 is western /k/ vs eastern /s/.
Is there a way to make things fit if we assume that one of the two groups actually retains the proto-sounds here, and derive the other group from that? Well… no, not really.
If the western group retained the original, we’d have to explain the difference between #1 and #3 in the eastern group, because the western group shows the same reflex, but the eastern group shows different reflexes. If the eastern group retained the original, we’d be in the same situation, except with #1 vs #2 instead of #1 vs #3.
Both have been attempted, but not successfully. There are enough words in each group to represent a very broad range of phonetic contexts, and there is simply no way to condition changes from one to the other that explains things adequately.
As it happens, we know it’s not that uncommon for palatals (like /c tʃ/) or palatalised velars (like /kʲ/) to become spirantised over time and end up as sibilants,2 so if we want to make some sort of sense of the correspondence between velars and sibilants in #3, we might want to consider the eastern sibilants as being the result of spirantisation from originally palatalised velars, so eastern is /kʲ/ instead of /s/. That means #3 is /k/ vs /kʲ ~ c/ – obviously better than /k/ vs /s/, but still two different sounds, so obviously not the proto-sound.
But wait – none of the western group actually have palatal(ised velar)s at all. What if #3 was actually just /kʲ ~ c/ everywhere, which was spirantised in the east and merged with plain /k/ in the west? Those are both well-attested sound changes, so that would be a reasonable hypothesis.
And what about #2, then? If we’ve now got rid of #3, it becomes possible to assume that the western group actually retains the original sound there, that is, a labiovelar /kʷ/. None of the languages in the eastern group have labiovelars at all, so a similar thing could have happened there: /kʷ/ and /k/ merged into /k/ (as indeed it did later on in the Goidelic branch of Celtic languages).
That would leave us with a three-way distinction in the proto-language, between /k/, /kʷ/ and /kʲ/. In the western group, /k/ and /kʲ/ merged while /kʷ/ remained distinct; in the eastern group, /k/ and /kʷ/ merged while /kʲ/ remained distinct (and was spirantised).
That means that in group #1, both groups have essentially retained the original (though with palatalisation before front vowels in the eastern group); in #2, the western group retained the original while the eastern group merged with #1; and in #3, neither group has retained the original, the western group merging with #1 and the eastern group spirantising (and in some cases merging with existing sibilants).
As it happens, this is one of the most well-known divides across the branches of Indo-European, and one of the first to be described. It is normally named by the Latin and Avestan words for ‘hundred’, which illustrate the split very neatly, and the western and easter groups are thus, respectively, called centum and satem languages. This is an isogloss which happens to divide the branches fairly neatly along an east-west line – with the notable exception of Tocharian, which is the easternmost Indo-European language discovered, but appears to have been mostly centum (‘western’).
It’s also important to note that the centum/satem divide is not phylogenetic – that is, languages in each group are not necessarily more closely related to each other internally than to languages in the opposite group. For example, it is commonly held that Greek and Armenian, though separate Indo-European branches, are closer to each other than to any other branches, even though Greek is centum and Armenian is satem.
Notes:
1 In actual fact, we don’t have to guess that this is what happened entirely: Mycenaean, the earliest form of Greek attested, uses a separate consonant to write this sound, neither k nor p, and it’s assumed that this actually reflects /kʷ/ directly.
2 This has famously happened in Romance languages (e.g. Latin centum /ˈkentũ(m)/ becomes French cent /sɑ̃/, Portuguese cento /ˈsẽtu/, etc.), and also in North Germanic (e.g., Old Norse kelda /ˈkelda/ becomes Swedish källa /ˈʃεlːa/).
