One of the main developments from Old Japanese to Middle Japanese is the reduction of several Ci/Cu syllables. Where old Japanese only had Cv syllables, Middle Japanese suddenly developed CvC, Cvv and CvN syllables. These seem to be generally accepted as rather haphazard, and irregular developments, as can be seen when Miyake (2003: 74) says "Owing to the extreme complexity of phonological change in post-Nara Japanese - particularly those changes called onbin ... - I will mostly deal with major, regular changes here. For more detailed discussion of onbin see Martin (1987), Frellesvig (1995), and Erickson (1998)".
Also Frellesvig in his fantastic 2010 overview "A History of the Japanese Language" presents onbin as primarily an irregular sound change whose general regularity cannot really be figured out anymore, and to a great extent is simply the result of these forms basically existing side by side for a long time. This work follows his 1995 work also cited by Miyake. I got that out of the library and find that I have some problems with the conclusions of Frellesvig. I think the development is presented as much more complex and irregular than it actually is, in a large part because of the fact that Frellesvig insists on seeing, what to me seem fairly clearly different developments, as a single development; Which subsequently makes it look like the shifts that took place could, in the same environment have two different outcomes. Once you actually allow for two different shifts a lot of nice attractive simplifications can be applied, and reflexes of Chinese loanwords can be integrated into the developments in an attractive way.
What are the onbin changes?
Apparently onbin has been used as a rather haphazard term for "euphonic" sound changes (which is quite literally what onbin means, so you can hardly blame people). But here we are specifically concerned with the developments that Frellesvig also looks at, namely, the reduction of *Ci1 and *Cu syllables yielding either a vocalic reflex or a consonantal reflex. These different reflexes, occasionally yield different outcomes in different dialects. Hence we get forms like:
*omopi-te "thinking" > Tokyo omotte (< *omoQte) Kyoto omoote (< *omoUte)
Other forms of verbs have a vocalic outcome in both:
*kaki-te "writing" > Tokyo/Kyoto kaite (< *kaIte)
When the reduced Cv syllable contained a pre-nasalized consonant (which in modern Japanese is generally just voiced: b, d, g) or a nasal consonant (m, n), the nasalization would be transferred to the following consonant:
*yobi-te"calling" > Tokyo yonde (< *yoNte) Kyoto yoode (< *yoŨte)
*sini-te"dying" > Tokyo/Kyoto sinde (< *siNte)
These developments are clearly most conspicuous in the verbal system where a suffix -te (and -tari, and in modern Japanese -ta) attached to the i stem triggers it, but they forms are found outside of the verbal system too. and there we find that the outcomes outside the verbal system are not always the same as the developments we find in the verbal system:
*taki-matu'torch' ('to burn' + 'pine) > taimatu
*kami-ta 'Familyname; god + rice paddy' > kanda
*aki-bito'tradesman' ('trade' + 'man') > akyuudo (< *akiŨto) ~ akindo (< (akiNto)
*ko-miti'backalley' ('small'+ 'road') > koodi (< *koŨti)
In a few cases an onbin syllable is dropped but nasality is transferred:
*kigisi > kizi 'Japanese pheasant' (< *kinsi < *kiNsi?)
*pi-mukasi > higasi'East' ('sun' + 'turning') (< *pinkasi < *piNkasi?)
*yama-miti > yamadi 'mountain-road') (< *yamanti < *yamaNti? cf. *ko-miti with a different resolution above!)
Three types of onbin outcomes
There are essentially three classes of onbin outcomes: Vocalic outcomes, Consonantal outcomes and pre-Nasalization outcomes. These three are all considered part of the same development by Frellesvig, and I think he is wrong to consider them part of the same process.
Pre-Nasalisation outcomes
pre-nasalization outcomes seem very marginal, and sometimes rely solely on internal reconstructions of the pre-Old Japanese form (as seems to be the case for those who take abiki 'trawling' as coming from *ami-piki'net pulling'); these are the types of outcomes we see in *kigisi, *pi-mukasi and *yama-miti. I do not know the corpus well enough to really evaluate the material, but it seems to me that a lot of these reductions go back to a Pre-Old Japanese phase, in which case they certainly cannot be part of the Middle Japanese onbin changes; and those that remain seem extremely marginal. Frellesvig in his 2010 book draws rather far-reaching conclusions from these outcomes. He compares *pumite 'brush' to *pumi-te 'stepping', which he considers to have been phonetically homophonous in the Middle Japanese period, but phonemically distinct, hence explaining why the former ended up as fude while the latter as funde. It is unclear to me whether *pumite is even attested in this form in Old Japanese, but I'm generally unhappy with words that are homophonous that end up with different outcomes. There is also almost certainly a difference in accent between these two words, which means they were probably never truly homophonous.
For now, I'm going to assume the pre-nasalization outcomes are marginal and are mostly the result of Pre-Middle Japanese developments, and will therefore be ignored here, as it doesn't seem obvious to me that they are part of the onbin changes we are looking at.
Consonantal outcomes
There are two type of consonantal outcome depending on whether the source syllable had a nasal or pre-nasalized consonant. If there was no nasal element the outcome is Q, i.e. an element that lengthens the following plain stop, e.g. omopi-te > omoQte> omotte.
The second type of consonantal outcome is the one that is triggered by a nasal or pre-nasalized consonant. Here this is written as N. It pre-nasalises the following consonant and inserts a moraic nasal n. Systematically these can basically be thought of as long pre-nasalised consonants, e.g. yomi-te 'reading' > yoNte> yonde.
While the consonantal outcomes are common in the verbal system, they are exceedingly rare in the outside of the verbal system, where the vocalic outcomes seem more common. This is one of the main reasons why I am inclined to see the vocalic and consonantal outcomes as the result of two subsequent shifts.
Vocalic outcomes
Vocalic outcomes come in two flavours, one that eventually ends up becoming an u and one that ends up becoming an i. Besides this, if the reduced consonant was pre-nasalized or a nasal the consonant following the u or i will be pre-nasalized.
The labials *pu, *pi; *bu, *bi; *mu, *mi always become U or Ũ; the vowel of the reduced syllable has no effect.
The velars however differentiate: I and Ĩ for ki and gi and U and Ũ for ku and gu
As a result, the I/Ĩ are a little rare.
Further Conditioning
As Frellesvig further shows, there are a whole bunch of distinguishing factors that block the onbin changes. First Old Japanese had two types of i vowels (but only before k, g, p, b and m) neutrally called i1 and i2. i2 is not part of the onbin changes. Alternations such as kami'god' and kamu-kaze'heavenly wind' have caused people to see i2 as the outcome of a contraction *uy. The fact that this diphthong can only stand in front of velars and labials suggests to me that the back-vowel has in fact transferred to labialization and we're rather dealking with kwi, pwi, mwi etc. It makes good sense that this would block the onbin changes. The distinction between these two i vowels is eventually lost in late middle Japanese, but only after the onbin changes had taken place.
Moreover, the development is blocked if the qualifying syllable is word-initial; and when the iCi1 and uCu sequences block the vocalic outcomes as well (but not the consonantal! this is important, because it solves an issue with the irregular verb iku'to go'). Also onbin of the labials does not take place before liquids (r) and glides (y, w). Frellesvig points out that the consonantal onbin of *pu/*pi only happens before *t.
My Solution
With this initial data out of the way, let me now get to my solution. I would argue for in fact two (or perhaps three) phases of reduction, first the vocalic onbin and later, the consonantal onbin. I have no idea if this suggestion has been made before; but I wanted and tried to figure it out myself, and I think this actually helps quite a bit. If anyone reads this that is much better read in the literature (which isn't exactly hard) and seems to recognise something someone has said before, do let me know!
Phase 1: Vocalic Onbin.
First, it seems to me that the vocalic onbin of the labials (p, b, m) took place, yielding U or Ũ in all cases. This shift is a shared innovation between all of Japanese (I wonder what Ryukyuan's position is in this; For some reason people who write about historical Japanese seem to generally ignore the one Japonic language that helps confirm some of the ideas...).
This gives rise to a whole bunch of long and complex vowels, which moreover gave rise to apparently nasalized vowels, which caused the next syllable to be pre-nasalized. In the verbal system, these appear to have been largely reversed analogically in the Tokyo dialect, while in Kyoto they were not. This then gives rise to forms such as: *kapite > Kyoto koot (but Tokyo katte); *yobi-te 'calling' > Dialectal yoode (no (longer?) in Kyoto) but not Tokyo yonde; *yomi-te 'reading' > Dialectal yoode (no (longer?)n in Kyoto) but not Tokyo yonde.
In fact, some of the *p final verbs stems in standard Japanese continue to have these onbin forms (in the same way that Kyoto has them regularly):
*top-u 'asks', *top-ite 'asking' > tou, toote not expected **totte
*kop-u 'begs', *kop-ite 'begging' > kou, koote not expected **kotte
*kop-u 'misses', *kop-ite 'missing' > kou, koote not expected **kotte (is this the same verb as 'begs'?)
While verbs of this type would have been affected in the forms before -t, hence yielding the non-Tokyo forms mentioned here,the *p, *b, *m appears to have been restored in the Tokyo dialect (and for Kyoto only the for *b and *m) in the bases of other stems of the verb where the sound was retained. There would also be some verbs that have the right shape to retain it in multiple parts of the verb, e.g. *tumu'to pile up, stack' and *tanosimu'enjoy' would not undergo the shift in more places of their paradigms than others.
Hence you'd find paradigms like:
tum-u 'piles up', tumi-te'piling up' tum-azu 'doesn't pile up' > tum-u, tuu-de, tum-azu
tanosim-u 'enjoys', tanosimi-te 'enjoying', tanosim-azu'doesn't enjoy' > tanosiu, tanosimi-te, tanosim-azu
Very marginal leftovers of this shift seem to be around still such as sumoo 'Sumo' which comes from sumaU < *sumapi, a deverbal/infinitival-like formation.
The volitional suffix *-amu has also undergone the onbin shift, and became *-aŨ yielding the modern Japanese -oo suffix.
Second, the vocalic onbin of the *k and *g takes place. This might have taken place at the same time as the labial one, but it's not completely clear. It seems really worth examining whether the shift was not more strictly conditioned than the labial one. For these verbs, the onbin form does not seem to have been analogically restored at all. One *kaki-te > kaIte > kaite; oyogi-te > oyoĨ-te > oyoide took place. Notice, however, that the shift does not take place in the final forms *kak-u > kaku not *kaU; idem for oyog-u > oyogu not **oyoŨ. Where they analogically restored, or did the shift simply never take place in word-final position in the Tokyo dialect? One is reminded of the same absence of the shift in the adverbial form of the adjective takaku does not become takaU, whereas takaki became takai (this is unlike the Kyoto dialect where both forms lose the *k).
Just as a quick aside: iki-te'going', which does not have the "expected" outcome iite in Tokyo Japanese, can be explained in this phase. as iCi sequences should block onbin, it makes perfect sense that it simply stayed iki-te; only later following the consonantal onbin. One wonders whether the form yuku, that is around every now and then is not a somewhat mixed form, but actually reflecting the original outcome of the final form *iku > iU. I do not know the answer to this, is yuku really the older form?
From the way that we see that Chinese words in *ŋ are borrowed, it seems clear that Ũ and Ĩ at some point had some kind of nasalized realization, and it's not just an abstract representation: *syaŋ'life' is borrowed as *syaŨ, eventually yielding syau > syoo, but with pre-nasalization of the next consonant in compounds, e.g. syoo-zuru'to produce' < *syaŨ-suru
It seems possible that, despite the eventual same outcome Ũ and Ĩ from *g and from *b originally had slightly different realizations, e.g. *gu > [ɣ̃ʷ] *gi > [ɣ̃ʲ] but *bu/bi/mu/mi > [β̃] or something along those lines.
Phase 2: Consonantal onbin
The first phase radically revamped the vocalic system drastically, introducing complex vowels and nasal vowels in a language that never had such things but only CV syllables.
The consonantal onbin, which Frellesvig took as the result of the same shift, but with a different outcome, is actually quite different from the previous shift. First of all, it affects several more consonants not involved with the vocalic one. The vocalic only affects the labials and velars; But the consonantal onbin affect t, r, n as well.
Moreover, the vocalic onbin has several factors that block the shift. First, identical vowels on both sides of the consonant; and Second, the i2 = [ʷi] is not elligible for onbin. Neither of these blocking factors seem to operate in consonantal onbin. These difference in consonants it affects and conditioning that takes place clearly suggests to me that we are dealing with an altogether different sound shift. Frellesvig is mistified by the fact that kabi2tati shifts to kandati; but this is perfectly understandable if you don't take these two shifts as being the same thing. Third, the consonantal onbin can not take place in word-final position.
The vast majority of the labial syllables that underwent the vocalic onbin are no longer available to undergo the next shift. Only newly built compounds, and the analogically restored forms in the verb forms will undergo consonantal onbin.
*p, *b, *m, *n shift to Q or N before other stops: *p, *t, *k and *n. I have not found any examples of it happening before m or s (unlike vocalic onbin where that can happen).
They yield verbal forms like: *omopi-te > omotte; *yobi-te > yonde; *yomi-te > yonde; *sini-te > sinde.
Examples of the shift elsewhere in the language can be found too:
With *m it is common in names with kami'god' or kami'up' in it:
kami-ta> kanda 'god field'
kami-kaki> kangaki'god fence'
kami-tati> kandachi'upper class'
Also at least one noun with kami in it has undergone this shift:
kami-nusi (or kamu-nusi) > kannusi'Shinto priest' ('god' + 'lord')
kami-tukasa (or kamu-tukasa) > kandukasa 'erson who administered religious rituals within the dazaifu jurisdiction (ritsuryo system)' (god + official)
The shortened form of nani, nan is probably also the result of this shift, e.g.:
nani-toki > nandoki 'when?' ('what' + 'time')
Other examples are:
umi-paru> funbaru'to brace one's legs' ('to step' + 'to stretch')
pumi-kiri> fungiri'determination' ('to step' + 'to cut')
pumi-tukeru> fundukeru'to trample on' ('to step' + 'to attach')
nomi-bee> nonbee'drunkard' ('to drink' + "bleh")
pumi-tukupe> fundukue'japanese desk' ('writing' + 'desk') also fudukue with irregular pre-nasalisation outcome.
kumi-dubokuretu> kundubokuretsu'locked in a grapple' (not sure how the word parses completely, but the first element is clearly from kumi'to grapple')
I have not yet found any example yet of the consonantal onbin of *p outside of the verbal system, nor have I found good examples of the vocalic onbin. Is this because it merged with the *w verbs, which perhaps do not regularly have onbin?
There are plenty of examples with *k, owing mostly to the fact that *piki is an intensifying verbal prefix:
*piki-paru > hipparu 'to pull tight' ('to pull'+ 'to stretch')
*piki-kosu > hikkosu 'to move' ('to pull' + 'to move')
*tuki-pasiru > tuppashiru 'to run swiftly' ('to stab' + 'to run')
*puki-tobu> futtobu'to blow off' ('blowing' + 'to fly')
*piki-tsuku> hittsuku'to stick to, adhere to' ('to pull' + 'to attach')
*piki-tsumaru> hittsumaru 'to snatch' ('to pull' + 'to grab')
*tuki-tuku> tsuttsuku'to poke repeatedly' ('to poke' reduplicated)
*piki-sageru > hissageru 'to carry in one's own hand' ('to pull' + 'to take by the hand')
*piki-nuku> hinnuku'to uproot' ('to pull' + 'to extract')
And hence as is very often the case with irregular verbs, the irregularity of iku'to go' is actual the result of following the regular sound laws: *iki-te > itte
According to Frellesvig, the development of *t, *r and *w having consonantal onbin is purely by analogy. All consonantal-final stems had onbin stems except for a few and hence onbin stems were spread to the rest. This fails to account for the fact that final *s- verbs never develop onbin stems. And while it seems possible that in the onbin-stems of the verbs the onbin forms could have spread by analogy, it's much more difficult to explain in compounds where no such onbin stem is in use otherwise. Yet we find many compounds that have reduction of *t and *r (I don't know about *w; it merges with *p, which definitely has onbin forms; hard to say what happens with original *w, a somewhat rare phoneme).
I've so far found hardly any examples of *t onbin forms where the following word does not start with a t. They may be conditioned to only take place in front of another t.
kati-te> katte'selfishness'
uttaeru'to raise, bring to someone's attention' must be a case of consonantal onbin too although I'm not sure about the etymology, probably uti-'to hit' + taeru'to support'? Any readers that know?
pitotu-pasiri> hitoppashiri 'giving something a whirl' (one + run)
For *r I've also found mostly forms in front of t, but also in front of some others:
kiri-te> kitte'stamp'
tori-te> totte'handle'
tori-tuki-nikuki> tottsukinikui 'difficult to approach' ('take' + 'attach' + 'difficult')
pori-tate> hottate'erection of a pillar by sinking it directly into the ground' ('to dig' + 'to stand up')
kusuguri-tai> kusuguttai'ticklish' ('to tickle' + deverbal adjectival suffix, cf. mede-ru 'to love, admire', mede-tai 'happy, auspicious').
siri-takaburi> sittakaburi'pretending to know' ('to know' + acting haughty')
nori-toru> nottoru'to take over' ('to ride' + 'to grab')
yari-tsukeru> yattsukeru'to attack an enemy' ('to do' + 'to attach')
sikameri-tura> shikamettsura'frown, grimace' ('to pull a wry face' + 'face')
tori-kumi-api > tokkumiai'grapple; scuffle' ('to hold'; 'to grapple'; 'to meet')
tori-parapi> topparai 'paying cash in hand' ('to hold'; 'to pay')
Conclusion
This interpretation, which assumes the vocalic onbin takes places before consonantal onbin, seems to account for the fact that words that shouldn't undergo the onbin changes, do undergo the consonantal one and not the vocalic one. The ordering of these rules as two seprate shifts really enlightens a lot of the problems. It is a little surprising to see how extremely many cases of compounds there are with the consonantal onbin and how few with the vocalic onbin. This seems to suggest that such compound verbs are fairly late formations. I do not know enough about Old Japanese to know whether it is true that compound verbs like this were somewhat rare.
The amount of nominal forms in general that clearly undergo the vocalic onbin is a little limited, and I would not be too opposed to the idea that most of those forms are the result of dialect mixing in the Tokyo dialect. *aki-bito > akyuudo being the Western outcome, and akindo being the regular Tokyo outcome. The only vocalic onbin shift which should almost certainly be considered to have taken place is the ki/gi > I/Ĩ and probably the *mu > Ũ shift in word-final position.