First of all, the number one consideration for whether rendaku happens or not is etymology: rendaku is common only in native Japanese vocabulary. It only very occasionally happens in words derived from Chinese (on-yomi) (株式会社 kabushiki-gaisha and 夫婦喧嘩 fūfu-genka, 口喧嘩 kuchi-genka, etc. probably being the most prominent exceptions), and is extremely rare in loans from other languages.Khemehekis wrote:Are there any correlations between a language having or lacking WH-fronting and other characteristics of said language?
Why do some Japanese compound words have rengaku and other Japanese compoind words don't?
The primary obstruction to rendaku in native vocabulary is a voiced obstruent in the latter morpheme (i.e. Lyman's law):
山 yama + 寺 tera = 山寺 yamadera
山 yama + 門 kado = 山門 Yamakado (not *Yamagado, because rendaku is blocked by /d/)
The reverse of Lyman's law can also block rendaku, i.e. when the last syllable of the first element contains a voiced obstruent, e.g.
島 shima + 田 ta = 島田 Shimada
永 naga + 田 ta = 永田 Nagata (not *Nagada, because rendaku is blocked by /g/)
Certain morphemes are more likely to force it or stop it. A prominent example is the pair of morphemes 大 ō "big" and 小 o "small". Probably due to their similar pronunciation but opposite meanings, rendaku is helpful in reinforcing the phonemic difference, so, e.g.
大 ō + 川 kawa = 大川 Ōkawa
小 o + 川 kawa = 小川 Ogawa
大 ō + 田 ta = 大田 Ōta
小 o + 田 ta = 小田 Oda
Some other morphemes just tend to resist rendaku all the time, e.g. 手 te:
山 yama + 手 te = Yamate (not *Yamade)
相 ai + 手 te = 相手 aite
苦 niga + 手 te = 苦手 nigate
片 kata + 手 te = 片手 katate
There are other things that can block it:
1. semantics: a dvandva compound ("X and Y") blocks rendaku:
山 yama + 川 kawa = 山川 yamakawa "mountains and rivers"
好き suki + 嫌い kirai = 好き嫌い sukikirai "likes and dislikes (related to food)"
山 yama + 川 kawa = 山川 yamagawa "mountain river"
食わず kuwazu + 嫌い kirai = 食わず嫌い kuwazugirai "dislike without ever tasting"
This is likely because the speaker in some way conceives of it more as a list than a single word.
2. branching constraints:
When the second of the two morphemes already belongs to a compound that the first one does not, rendaku is blocked:
紋 mon + (白 shiro + 蝶 chō) = 紋白蝶 monshirochō (family crest + (white + butterfly))
本 hon + (駒 koma + 込め kome) = 本駒込 Hon-Komagome (true (Komagome (placename)))
(尾 o + 白 shiro) + 鷲 washi = 尾白鷲 ojirowashi ((white + tail) eagle)
本 hon + 腰 koshi = 本腰 hongoshi (true (lower.back))
Just for your information, the Wikipedia article gives 山火事 yamakaji (not *yamagaji) as an example of Lyman's law: supposedly kaji does not become *gaji because it is blocked by /ʥ/. This is actually a very bad example because first of all, kaji is Sino-Japanese vocabulary and thus unlikely to undergo rendaku under any case. Secondly, branching constraints forbid *yamagaji because kaji is already a compound (yama + (ka + ji)). I should probably edit the article to use Yamakado as in my example here.
Despite all this, rendaku is not totally predictable. It is subject to regional variation and other factors. Place names and surnames in particular can be confusing. Some surnames can be read either way, e.g. 山崎 can be Yamazaki or Yamasaki, while some place names have unexpected rendaku (長渕 Nagabuchi, not expected Nagafuchi) or lack expected rendaku (中島 Nakashima, not Nakajima), or have multiple possibilities. 秋葉原 Akihabara (aki + ha + hara), for example, if it were not so well-known, would be a confusing place name, as it could also be read Akibahara (in fact the nickname of Akihabara is Akiba).
For the word くらい/ぐらい kurai/gurai, it seems to be totally optional if you want to use it or not, though it seems like gurai is more colloquial and kurai more formal.
Side note: It's also interesting that branching constraints affect gemination. For example, when the first element of a compound ends in -ku and the second element begins with /k/, they will merge to form -kk. Additionally, when the first element ends in -tsu and the second element begins with any unvoiced consonant, it will form a geminate.
学 gaku + 校 kō = 学校 gakkō.
発 hatsu + 車 sha = 発車 hassha
立 ritsu + 体 tai = 立体 rittai
切 setsu + 腹 fuku = 切腹 seppuku
However, the same type of branching constraint will disallow these, but with an interesting difference between the two:
-tsu is always disallowed, e.g.
(生 sei + 活 katsu) + 費 = 生活費 seikatsuhi "cost of living" (not *seikappi)
年 nen + 末 matsu + 調 chō 整 sei = 年末調整 nenmatsu-chōsei "year-end tax adjustment" (not *nenmatchōsei)
The same restriction applies for -ku, but colloquially this rule is entirely ignored:
(洗 sen + 濯 taku) + 機 ki "washing machine" = 洗濯機 sentakuki officially, but sentakki colloquially
(水 sui + 族 zoku)j + 館 kan "(public) aquarium" = 水族館 suizokukan officially, but suizokkan colloquially
EDIT: I found a very interesting looking pdf about this (it's in Japanese though) that goes into a lot of detail. I'd like to read it and share any findings it may have when I have more time.