1 Introduction
Lyman's Law was first proposed as a constraint on a well-known morphophonemic phenomenon in modern Tokyo (‘standard’) Japanese called rendaku (see §2), which yields alternations such as /k/ ~ /g/, as in /kame/ ‘turtle’ and /umi+game/ ‘sea turtle’ (cf. /umi/ ‘sea’). Lyman's Law prohibits rendaku from occurring in the second element of a compound if that element already contains a medial voiced obstruent. Consequently, we find compounds like /ita+kabe/ ‘wooden wall’ (cf. /ita/ ‘board’, /kabe/ ‘wall’), in which the medial /b/ in the second element prevents the form with rendaku: */ita+gabe/.Footnote 1
We argue that Lyman's Law originated as a prohibition against prenasalisation in consecutive syllables, and that this prohibition is what held in 8th-century Old Japanese. Support for this claim comes partly from the phonetics of Old Japanese, which we consider briefly in the following paragraph, and more thoroughly in §3 and §4. Additional support comes from modern dialects of Japanese that retain both prenasalisation and the original prohibition (§5), as well as from philological analyses of Old Japanese compounds, which are entirely consistent with a prohibition against prenasalisation in adjacent syllables (§6.1).
The earliest substantial written documents for Japanese record varieties of Old Japanese. At that time, the consonants corresponding to modern Tokyo /b d ʣ g/ were pronounced [ᵐb ⁿd ⁿz ᵑg] and, we assume, triggered salient nasalisation on an immediately preceding vowel (see §3.2).Footnote 2 The evidence for prenasalisation is overwhelming, but even a cursory review would take us far beyond the scope of this paper. We also find prenasalised voiced obstruents in some highly stigmatised modern dialects of Japanese (§5), and this prenasalisation is generally regarded as a retention from Old Japanese times.
Rendaku was already well established in Old Japanese, but, as noted above, the Old Japanese precursor of Lyman's Law seems to have been a straightforward ban on prenasalised voiced obstruents in adjacent syllables (§3.4). Stanton (Reference Stanton2019) thoroughly documents a phonetically natural and typologically common restriction against /NCVNC/ sequences (where /N/ is a nasal consonant and /C/ an obstruent), and the Old Japanese version of Lyman's Law follows from this restriction, provided that prenasalised obstruents are like /NC/ clusters in the relevant respects (§4).
The details of the transition to the modern Tokyo version of Lyman's Law are not well understood. In Old Japanese, rendaku yielded a prenasalised voiced obstruent at the beginning of the second element of a compound (§3.3), as in OJ /asi+ᵑgamo/ ‘reed duck’ (cf. /asi/ ‘reed’, /kamo/ ‘duck’), but the constraint on adjacent syllables blocked rendaku when there was a prenasalised voiced obstruent in the syllable immediately preceding or immediately following a potential rendaku site (§3.4). In modern Tokyo Japanese, a voiced obstruent anywhere in the second element of a compound prevents rendaku; the inhibitor consonant need not be in the syllable immediately following a potential rendaku site (§2). Furthermore, a voiced obstruent preceding a potential rendaku site has no effect; even if it is in the immediately preceding syllable, it does not block rendaku (§3.4). Lyman's Law has thus undergone a metamorphosis from a phonetically natural restriction on adjacent syllables into a phonetically unnatural and cross-linguistically rare long-distance restriction on voicing.
2 Rendaku and Lyman's Law in modern Tokyo Japanese
Many modern Tokyo morphemes have one allomorph that begins with a voiceless obstruent and another allomorph that begins with a voiced obstruent, as in (1).
-
(1)
The initial voiced obstruent in the /dake/ allomorph in (1c) is an instance of rendaku, and the prototypical environment for rendaku is the beginning of the second element of a two-element compound. There are, however, some instances of rendaku in the second elements of words that can be analysed as prefix + base derivatives rather than as compounds. Also, many compounds have more than two elements, and some instances of rendaku are in third or later elements. Only the non-rendaku allomorph of an alternating morpheme can occur word-initially, such as /take/ in (1a) and (b).
A few words have been coined by exploiting a well-known phonaesthetic association between voiced obstruents and mostly negative attributes, particularly in initial position in native words (Suzuki Reference Suzuki1962: 23–24, Endō Reference Endō1977: 222–228, Komatsu Reference Komatsu1981: 87–88). One example is /ʣama/ ‘sorry state’, which is obviously related to /sama/ ‘state, condition’. Since /ʣama/ and /sama/ differ both in pronunciation and in meaning, they cannot be analysed as allomorphs of the same morpheme; they are separate lexical items, and the /ʣ/ in /ʣama/ is not an instance of rendaku. A word like /ʣama/ can be treated as a derivative, marked by replacing a voiceless obstruent with its voiced partner (Suzuki Reference Suzuki1962: 26–27), although this analysis is not always appropriate synchronically, since native speakers do not see the connection in every such pair. The subsegmental exponent of this derivational morpheme just happens to be homophonous with the exponent of rendaku in an analysis that treats rendaku as the manifestation of a subsegmental linking morpheme joining elements of a compound (as in §4). We should also note here that, in a few cases, a rendaku allomorph has ousted its non-rendaku counterpart diachronically. For example, the ancestor of the second element in /migi+gawa/ ‘right side’ was /kawa/, but most speakers today pronounce the independent word meaning ‘side’ as /gawa/. Once a morpheme ceases to alternate in this way, its initial voiced obstruent is no longer a synchronic instance of rendaku.
Since the 1960s, rendaku has attracted a great deal of attention from theoretical phonologists. Although rendaku always pairs a voiceless obstruent with a voiced obstruent, the phonetic difference between the two phonemes in each pair is in many cases more than just the absence vs. presence of voicing (Vance Reference Vance2016: 3–4). The examples in (2) show the pairings. All the modern Tokyo obstruent phonemes except /p/ are involved in the alternations. Each voiceless obstruent phoneme other than /p/ (i.e. /ɸ h t ʨ ɕ ʦ s k/) appears on the left in at least one of the pairings, and each voiced obstruent (i.e. /b d ʣ ʥ g/) appears on the right in at least one of the pairings. Notice that /ɸ/ [ɸ] and /h/ [ç, h] both alternate with /b/ [b], /ʨ/ [ʨ] and /ɕ/ [ɕ] both alternate with /ʥ/ [ʥ], and /ʦ/ [ʦ] and /s/ [s] both alternate with /ʣ/ [z, ʣ].
-
(2)
The historical changes that led to the pairings in (2) are well understood (Vance Reference Vance2015: 397–398), and the pairings before these changes began will be described below in §3.3. For the minority of Tokyo speakers who maintain a conservative variety with word-medial syllable-initial [ŋ] (Hibiya Reference Hibiya and Tsujimura1999), the pairing in (2i) is [k] ~ [ŋ], not [k] ~ [g]. This additional deviation from straightforward [±voice] will be relevant below in §4.Footnote 3
The constraint known as Lyman's Law (Vance Reference Vance2015: 402–408) prevents rendaku in a two-element compound if the second element (E2) contains a medial voiced obstruent. The examples in (3) illustrate.
-
(3)
There are very few exceptions to Lyman's Law in the existing vocabulary (Kindaichi Reference Kindaichi1976: 5, Martin Reference Martin1987: 115, Suzuki Reference Suzuki2005). The best-known exception is probably /nawa+baɕigo/ ‘rope ladder’ (cf. /nawa/ ‘rope’, /haɕigo/ ‘ladder’). It does not seem to matter whether the inhibiting voiced obstruent is in the syllable immediately following the potential rendaku site or later in E2 (Lyman Reference Lyman1894: 161–162, Martin Reference Martin1952: 48). The examples in (4) are typical.
-
(4)
The results of some psycholinguistic experiments (Vance Reference Vance1980: 258–259, Ihara & Murata Reference Ihara and Murata2006: 21–22) have been consistent with a small ‘distance effect’ in responses to compounds containing nonsense E2s: the more syllables there were between the potential rendaku site and the inhibitor voiced obstruent, the more likely participants were to choose responses with rendaku. A more recent study by Kawahara (Reference Kawahara2012), however, using a different methodology, found no such effect. The apparent distance effect in the earlier studies was probably due to some other variable that was uncontrolled (Vance Reference Vance2015: 403–404).
The term ‘rendaku-eligible’ can be used to refer to an element that begins with a voiceless obstruent as an independent word. Words that begin with a vowel, with a sonorant consonant or with a voiced obstruent, like the examples in (5), are not rendaku-eligible elements, because rendaku cannot apply.
-
(5)
As Kawahara (Reference Kawahara2016: 33) notes, some accounts in the literature can easily be misinterpreted to mean that rendaku always applies to rendaku-eligible E2s unless Lyman's Law would be violated. In fact, however, rendaku is intractably irregular. Many phonological, morphological and semantic factors are known to influence the likelihood of rendaku, but there is no overarching generalisation that predicts when it occurs and when it does not.
First, some rendaku-eligible elements are idiosyncratically immune (Irwin Reference Irwin2016: 104–105), including /kasu/ and /himo/, as shown in (6).
-
(6)
The allomorphs */gasu/ and */bimo/ simply do not exist, despite the fact that neither contains a medial voiced obstruent. Here, and throughout, we mark all non-occurring forms with an asterisk, regardless of whether there is any principled explanation for their non-occurrence. That is, a form marked with an asterisk is not necessarily ‘ungrammatical’ in any meaningful sense.
Second, many rendaku-eligible elements behave inconsistently as E2s, appearing sometimes with rendaku but sometimes without, even when no putative inhibiting factor is relevant. The examples in (7) illustrate.
-
(7)
Both E2s in the compounds in (7) are polysemous (as almost all content morphemes are), but the compounds were chosen so that E2 in each pair of compounds would have the same sense. This precaution is necessary because the probability of rendaku in an E2 can differ dramatically depending on its sense (Vance Reference Vance2015: 433, Irwin Reference Irwin2016: 104–105).
Third, there are individual compounds, such as those in (8), that can be pronounced either with or without rendaku.
-
(8)
There are many well-known examples of compounds gaining or losing rendaku over time (Vance Reference Vance, Frellesvig, Shibatani and Smith2007a: 163), and variability like that in (8) is often symptomatic of a shift in progress.
Despite the pervasive irregularity of rendaku, Lyman's Law is a very robust constraint. It is only natural that the interaction between this constraint and the rendaku alternations has attracted the attention of many phonologists (e.g. Kawahara & Zamma Reference Kawahara and Zamma2016: 18–26). The primary aim of this paper is to shed light on the nature of Lyman's Law by investigating its historical origin. In §3 we look at Old Japanese phonology, and in §4 we show that widely accepted inferences about Old Japanese phonetics motivate an origin scenario involving a prohibition against prenasalisation in adjacent syllables. In §5 we look at non-standard modern dialects of Japanese that retain prenasalisation and the original prohibition. In §6 we show that the philological evidence is consistent with our claim that the Old Japanese forerunner of Lyman's Law was a constraint on adjacent syllables, although some inflectional forms and lexicalised phrases present a challenge. We offer a plausible explanation for these apparent exceptions in §6.2.
3 Old Japanese and the strong version of Lyman's Law
3.1 Overview
To investigate the historical origin of Lyman's Law, we first review the basic phonological structure of Old Japanese in §3.2, focusing on prenasalised voiced obstruents. We then present a plausible account of how rendaku originated in §3.3. Finally, in §3.4 we examine the Old Japanese precursor of Lyman's Law (often called the ‘strong version’ of Lyman's Law).
3.2 Prenasalised obstruents in Old Japanese
As noted in §1, rendaku was already well established in Old Japanese, but voiced obstruents were prenasalised. The phoneme chart in (9) displays the consonant system that most historical linguists infer for Old Japanese (Miyake Reference Miyake2003: 74, Frellesvig Reference Frellesvig2010: 34–36, Bentley Reference Bentley and Tranter2012: 191).
-
(9)
The chart in (9) makes Old Japanese look like a language in which prenasalisation could be characterised as a voicing enhancement, since there is no series of plain voiced obstruents (Riehl & Cohn Reference Riehl and Cohn2011: 554). There are good reasons to believe, however, that the Old Japanese ‘voiceless’ obstruents were allophonically voiced word-medially, i.e. intervocalically (Frellesvig Reference Frellesvig2010: 35). It follows that prenasalisation was the distinctive feature, not just an enhancement, for word-medial ‘voiced’ obstruents. Understandably, many historical linguists avoid describing the prenasalised obstruents of Old Japanese as voiced and the contrasting obstruent series as voiceless. Frellesvig (Reference Frellesvig2010: 34–36) adopts the terms ‘media’ (for the prenasalised series) and ‘tenuis’ (for the non-prenasalised series). Other scholars prefer ‘lax’ and ‘tense’.
Old Japanese prenasalised obstruents occurred only word-medially in the non-mimetic native vocabulary (Frellesvig Reference Frellesvig2010: 43, Takayama Reference Takayama2015: 627–629). The usual diachronic explanation for this phonotactic restriction attributes prenasalised obstruents to prehistoric NC clusters that originated as contracted NVC sequences (Ramsey & Unger Reference Ramsey and Unger1972: 278, Miyake Reference Miyake2003: 73, Frellesvig Reference Frellesvig2010: 42–43). It appears that such contraction was not possible in word-initial NV syllables. Since Old Japanese did not allow coda consonants, it is not really necessary to assume that the inferred contraction process produced phonological consonant clusters in prehistoric Japanese. Consider the example in (10).
-
(10)
Assuming that intermediate phonetic forms like [muɾãnsi] actually occurred, they could have been fast-speech variants whose phonemic forms still included the elided vowel. But once voicing assimilation took place (yielding [nz] or [ⁿz]), reanalysis as a prenasalised consonant would have been compelling for later generations of speakers, assuming the prohibition against closed syllables remained active. Since prenasalisation seems to be unstable in word-initial position cross-linguistically (Herbert Reference Herbert1986: 18), it could be that even word-initial NV contraction was phonetically possible in prehistoric Japanese, but that postnasal voicing and reanalysis were blocked.
On the other hand, the scenario in (10) depends crucially on the anticipatory nasalisation of the vowel preceding the contracted NV syllable. Cross-linguistically, such regressive assimilation is typically stronger and more likely to be marked in a phonetic transcription when the immediately following nasal consonant is in the coda of the same syllable rather than in the onset of the next syllable (Herbert Reference Herbert1986: 126, Stanton Reference Stanton2019: 658). This generalisation is what motivates the [ã] assumed in (10) for contracted [mu.ɾãn.si], but not for uncontracted [mu.ɾa.nu.si]. The transcription [ã] also appears in reanalysed [muɾãⁿzi], because vowel nasalisation is typical immediately preceding a prenasalised obstruent, but not immediately preceding a simple nasal (Herbert Reference Herbert1986: 134). This difference is hard to explain if the relevant prenasalised obstruents and ordinary nasals are all simple onset consonants, and it is one of the reasons that Herbert (Reference Herbert1986: 134) gives for analysing prenasalised obstruents as contour segments which are ‘ambisyllabic’ in the sense that the nasal portion affiliates with the syllable to its left, while the oral portion affiliates with the syllable to its right.
Needless to say, we cannot prove that the Old Japanese nasalisation facts were as we describe them here in §3.2. According to Jeong (Reference Jeong2012: 450), some languages seem to have strong anticipatory nasalisation immediately preceding an onset nasal consonant. If this was true in Old Japanese, it would undermine our argument that Old Japanese prenasalised obstruents were ambisyllabic in Herbert's sense. Jeong also reports that some languages seem to have relatively weak anticipatory nasalisation immediately preceding a coda nasal consonant, although stronger than immediately preceding an onset nasal consonant. If the nasalisation of the first vowel was weak both in Old Japanese ṼNV and in Old Japanese ṼNCV and too similar to discriminate reliably, this would also undermine our argument.
3.3 The origin of rendaku
The scenario in §3.2 for the historical development of prenasalised obstruents dovetails with a widely accepted account of the origin of rendaku. In contrast to the modern Tokyo phoneme pairings shown by the examples above in (2) in §2, the Old Japanese pairings, shown below in (11), were phonetically uniform.
-
(11)
In a typical Old Japanese noun + noun compound noun with rendaku, such as OJ /matu+ᵐbara/ ‘pine grove’ (cf. /matu/ ‘pine’, /para/ ‘field’), it is reasonable to suspect that prenasalised /ᵐb/ developed from a prehistoric sequence of the form /NVp/. The obvious candidate for the NV syllable here is the ancestor of the Old Japanese genitive particle /no/ (cf. modern Tokyo /no/) (Murayama Reference Murayama1954: 107, Unger Reference Unger1975: 8–9, Vance Reference Vance1982: 335–338, Frellesvig Reference Frellesvig2010: 40–43), as in (12).Footnote 4
-
(12)
Like Pre-OJ /mura+nusi/ in (10), the prehistoric form in (12) is, of course, hypothetical. The authoritative dictionary of Old Japanese (Jōdaigo Jiten Henshū Iinkai 1967) lists the compound /matu+ᵐbara/, as well as its two elements as headwords, but not a frozen phrase in which the two elements are linked by the genitive particle. The phonographic attestations of this compound and of those in (13) below are all unambiguous, in the sense explained below in §4. Prenasalisation is transcribed phonetically throughout this paper as nasalisation on the immediately preceding vowel, followed by a short nasal with oral closure (see §3.2 above), but actual realisations in the present-day Kahoku-chō dialect (see §5 below) sometimes lack a measurable oral closure (Miyashita et al. Reference Miyashita, Irwin, Wilson and Vance2016: 186). Herbert (Reference Herbert1986), Maddieson & Ladefoged (Reference Maddieson, Ladefoged, Huffman and Krakow1993) and Riehl & Cohn (Reference Riehl and Cohn2011) provide cross-linguistic surveys of the phonetics and phonology of prenasalised consonants.
There is, however, no reason to assume that every Old Japanese noun + noun compound noun developed from an ancestor of the form Pre-OJ noun + /no/ + noun. Modern Tokyo Japanese has frozen noun + /no/ + noun phrases like /te+no+hira/ ‘palm of the hand’ (cf. /te/ ‘hand’, /hira/ ‘flat’) as well as noun + noun compounds without rendaku like /te+kase/ ‘hand shackles’ (cf. /kase/ ‘shackles’). The situation in prehistoric Japanese was probably much the same. In Old Japanese, as expected, rendaku did not occur in frozen phrases (aside from one puzzling exception; Vance Reference Vance, Frellesvig, Shibatani and Smith2007a: 164), and there are also many compounds without rendaku (presumably created by simple noun + noun juxtaposition). The examples in (13) illustrate; the first element in both is OJ /ko/ (~ /kwi/) ‘tree, wood’.
-
(13)
Assuming that the prehistoric ancestors of the examples in (13), when they were first coined, were essentially identical to their Old Japanese forms, they are entirely compatible with the account proposed above for the origin of rendaku: (13a) contains an NV particle that did not contract, and (13b) never had an NV particle in the first place. These examples also show why rendaku was intractably irregular in Old Japanese, just as it continues to be in modern Japanese (as noted above in §2): both noun + noun combinations (as in (13b)) and noun + genitive + noun combinations (as in (12)) were possible.
The irregularities have never been levelled out, and sporadic changes continue to this day, with some vocabulary items gaining rendaku and others losing it, as noted above in §2. Labrune (Reference Labrune2016) surveys irregularly occurring compound markers in several languages; many of these markers seem to have originated historically as contracted genitive markers. As Frellesvig (Reference Frellesvig2010: 40–41) points out, there are examples of rendaku in Old Japanese that do not seem to be derivable from any earlier phrase with an NV syllable between the elements, and he draws the reasonable conclusion that ‘rendaku already in OJ was established as a morphophonemic process’ that could trigger analogical extensions.
3.4 Lyman's Law in Old Japanese
The precursor of Lyman's Law in Old Japanese was markedly different from the modern Tokyo version described above in §2. It appears that rendaku was blocked if there was a (prenasalised) voiced obstruent in an adjacent syllable on either side of the target segment. This stricter constraint is known as the ‘strong version’ of Lyman's Law (Ramsey & Unger Reference Ramsey and Unger1972: 287–289). The examples in (14) illustrate.
-
(14)
There was no (prenasalised) voiced obstruent in OJ /tama/ (14a) or in OJ /pune/ (b), but there was in the last syllable of OJ /apaᵐbi/ (14a) and in the last syllable of OJ /suⁿzu/ (b). Neither E2 was idiosyncratically immune to rendaku, as shown by (14c) and (d).
The Old Japanese examples cited above are all attested phonographically, which means that they appear in Old Japanese texts represented in man'yōgana, i.e. Chinese characters used as phonograms (mostly syllabograms). OJ /suⁿzu+pune/ in (14b) appears only once, written 須儒赴泥, with each character representing a syllable. This sole occurrence of OJ /suzu+pune/ is in poem 51 in Nihon shoki (720).Footnote 5 The character 儒 unambiguously represents OJ /ⁿzu/, and the character 赴 unambiguously represents OJ /pu/ (see §4 and §6.2 below for more details on ambiguous phonogram spellings). The three logographic occurrences of OJ /apaᵐbi+tama/ (14a) are written 鰒玉 or 鰒珠, with each character representing a morpheme. The two phonographic occurrences (poems 4101 and 4103) are both in book 18 of Man'yōshū (ca. 760), and both are written 安波妣多麻. The character 妣 unambiguously represents OJ /ᵐbi/, and 多 unambiguously represents OJ /ta/. It is, of course, inappropriate to use the inferred pronunciation of a logographic attestation as evidence in phonological analysis. In particular, there is no direct evidence for the presence or absence of rendaku in a logographically written compound. At the same time, a phonographic representation of a compound is not necessarily an accurate reflection of that compound's actual phonemic form. No existing Old Japanese texts are original; they are handwritten copies of handwritten copies, sometimes with discrepancies between different copies (Frellesvig Reference Frellesvig2010: 22), and it is implausible to imagine that the critical editions on which scholars rely are entirely free of transmission errors. Furthermore, some individual man'yōgana characters were used inconsistently, sometimes representing a syllable beginning with a ‘voiceless’ obstruent and sometimes representing a syllable beginning with the corresponding (prenasalised) voiced obstruent. In addition, for some compounds attested phonographically more than once, it appears that forms with and without rendaku coexisted. Consequently, for a compound attested only once, there is no way to know for sure whether or not it was variable with respect to rendaku. Historical linguists working on Old Japanese simply have to keep these philological facts of life in mind, and do the best they can with the material that is available. Several other phonographically attested Old Japanese compounds will be cited below; these examples all need to be interpreted with caution. We report carefully on inconsistently used phonograms below in §4 and in §6.2, where they are directly relevant to our arguments, but there is no getting around the possibility that an unambiguous phonogram could simply be an error.
The strong version of Lyman's Law clearly does not hold in modern Tokyo Japanese. There are many examples like those in (15), in which rendaku occurs even though the last syllable in E1 contains a voiced obstruent. Since this voiced obstruent is not in the same element as the rendaku site, the modern version of Lyman's Law (§2) does not apply.
-
(15)
The results of a psycholinguistic study (Kawahara & Sano Reference Kawahara and Sano2014) indicate that the strong version of Lyman's Law is not psychologically real for modern speakers.
4 Lyman's Law and the Obligatory Contour Principle
We described the Old Japanese strong version of Lyman's Law just above in §3.4 as a ban on prenasalisation in adjacent syllables. In this section, we construe this ban as an instance of a cross-linguistically common dispreference for /NCVNC/ sequences (where N is a nasal stop and C is an obstruent). As Stanton (Reference Sanders2019) demonstrates, perceptual confusion provides a clear phonetic basis for disfavouring such sequences. The version of Lyman's Law that we see in modern Tokyo Japanese prevents multiple instances of voicing within a morph, and the relevant syllables need not be adjacent (§2). Such a constraint on voicing is cross-linguistically rare, presumably because it lacks phonetic motivation, as we will argue below.
Ever since an influential article by Itô & Mester (Reference Itô and Mester1986), it has been popular to interpret Lyman's Law as a manifestation of the Obligatory Contour Principle (OCP). Originally proposed as a prohibition against sequences of identical tones in underlying representations (Leben Reference Leben2011: 326), the OCP was incorporated into autosegmental phonology as a constraint on the tonal tier (Goldsmith Reference Goldsmith1990: 309–318). The notion of putting non-tonal features on separate tiers made it possible to treat such features as effectively suprasegmental, and a ban on multiple occurrences of voicing could then be formulated as a subcase of the OCP: OCP[voice] (Odden Reference Odden, Goldsmith, Riggle and Yu2011: 22).
To make this approach work for Lyman's Law, voicing must be a monovalent feature applicable only to obstruents, the sole class of segments for which voicing is distinctive in Japanese. Itô & Mester (Reference Itô and Mester1986), Mester & Itô (Reference Mester and Itô1989) and Rice (Reference Rice1993) offer different theoretical implementations of this idea. If the non-distinctive voicing of vowels and sonorant consonants were also specified, any morpheme with two or more voiced segments in a row, such as /kiʣu/ ‘wound’ or /ame/ ‘rain’, would violate OCP[voice]. If voicing were treated as a traditional binary feature that specifies voiced obstruents as [+voice] and voiceless obstruents as [―voice], the OCP would not block rendaku in compounds like /ko+hiʦuʥi/ in (4b), because the [+voice] specifications in */ko+biʦuʥi/ associated with /b/ and /ʥ/ would be separated by the [―voice] specification associated with /ʦ/.
Rendaku itself can be treated as a subsegmental linking morpheme that joins the two elements of a compound (Itô & Mester Reference Itô and Mester1986: 56–57, Vance Reference Vance2015: 406–407, Labrune Reference Labrune2016). This morpheme has to be something more abstract than just the phonetic feature [voice], of course, since it converts a voiceless obstruent into its rendaku partner, as in (2) above (Vance Reference Vance, Fukuda, Kim and Park2018: 193–197). The linking element must also be prevented from docking onto anything other than an obstruent at the beginning of the second element of a compound (Vance Reference Vance2015: 406). Otherwise, it could produce forms like */jugi+jama/ instead of /juki+jama/ ‘snowy mountain’ (cf. /juki/ ‘snow’, /jama/ ‘mountain’) or */ura+maʥi/ instead of /ura+maʨi/ ‘backstreet district’ (cf. /ura/ ‘rear’, /maʨi/ ‘town’).
Given these assumptions, Lyman's Law works as shown in (16). For the sake of illustration, the autosegmental representations are simplified to show a separate voicing tier, with all other segmental information consolidated into a single tier. OCP[voice] blocks rendaku in /ao+kabi/ ‘blue-green mould’ in (16a) (cf. /ao/ ‘blue, green’, /kabi/ ‘mould’), but not in /iro+gami/ ‘coloured paper’ in (b) (cf. /iro/ ‘colour’, /kami/ ‘paper’).
-
(16)
Since there are two adjacent [voice] specifications in (16a), OCP[voice] erases the unlinked one, or at least prevents it from linking, ensuring that the compound does not surface as */ao+gabi/.
The domain of OCP[voice] has to be limited, of course, so that it applies only to voicing specifications that would otherwise end up linked to segments in the same morph. Adjacency across a morpheme boundary does not count. Without this limitation, any word containing more than one voiced obstruent would be a violation, including the examples in (15) in §3.4 and even examples like /abura+ʦubo/ ‘oil pot’.
A problem arises, however, when we consider Sino-Japanese binoms, which are written with two Chinese characters, and are at least arguably bimorphemic. A minority of Sino-Japanese binoms can undergo rendaku as E2s, including /ka+ɕi/ ‘sweets’, as in /wata+ga+ɕi/ ‘cotton candy’. However, no Sino-Japanese binom with a medial voiced obstruent ever undergoes rendaku, despite the fact that the medial voiced obstruent is never in the same morph as the target consonant for rendaku (Vance & Asai Reference Vance and Asai2016: 121). There are no examples like */jama+ga+ʥi/, as opposed to actually occurrinɡ /jama+ka+ʥi/ ‘mountain fire’ (cf. /ka+ʥi/ ‘fire incident’). Thus Sino-Japanese binom elements behave like native monomorphemic elements. On the other hand, if Sino-Japanese binoms are treated as if they are monomorphemic, they can violate OCP[voice] with impunity, as in /bo+go/ ‘native language’, /dai+gaku/ ‘university’ and many other words. Incidentally, recent monomorphemic loanwords are exempt from OCP[voice], as in /adobaisu/ ‘advice’. The usual assumption, therefore, is that OCP[voice] applies only to native morphemes (Nasu Reference Nasu2015: 258–259).
Notice that the strong version of Lyman's Law in Old Japanese (§3.4) was not just the present-day version with a larger domain (Vance Reference Vance, Mufwene, Francis and Wheeler2005: 36–37). That is, the strong version was not simply a matter of limiting (prenasalised) voiced obstruents to one per word as opposed to one per morph, because adjacency was relevant.Footnote 6 As noted below in (24) in §6.1, the two voiced obstruents in OJ /joroⁿdu+taᵐbi/ ‘many times’ and in reduplicated OJ /tuᵑgi+tuᵑgi/ ‘again and again’ are non-adjacent, so neither violates the strong version of Lyman's Law. Both would violate a word-domain version of OCP[prenasal], however, despite the fact that neither has rendaku.
In conservative varieties of modern Tokyo Japanese that have word-medial [ŋ] (<OJ [ᵑg]) instead of [g], [ŋ] blocks rendaku, just as a voiced obstruent does (Kawahara & Zamma Reference Kawahara and Zamma2016: 25–26). The examples in (17) illustrate.
-
(17)
E2s with medial [m] (17a) or [n] (17b) do not block rendaku. Along with the phonetically opaque pairings of voiceless and voiced obstruents in (2) (§2), the inhibiting power of velar nasals shows that Lyman's Law in modern Tokyo Japanese has lost its phonetic grounding. It presumably has to be learned as a language-specific constraint.
The question of whether syllable-initial [ŋ] is an allophone of /g/ or a separate phoneme is controversial (Vance Reference Vance2008: 214–222) and will not be addressed here. Native speakers of Tokyo Japanese who do not have [ŋ] tend to assume that [ŋ] and [g] must be allophones of the same phoneme. On the assumption that the relationship between [ŋ] and [g] is allophonic, the interaction with Lyman's Law leads to an opacity problem in an OT analysis. This assumption may well be wrong, but Ito & Mester (Reference Ito, Mester, Fanselow and Féry2003b) accept it, and as a result, they have to resort to a stratal version of OT. For discussions of opacity, sound change, and phonetically unnatural patterns, see Sanders (Reference Sanders2003), Hayes & White (Reference Hayes and White2015) and Beguš (Reference Beguš2020).
Furthermore, Kawahara (Reference Kawahara2008: 324–327) has argued that OCP[voice] does not seem like a plausible universal constraint in the first place. Cross-linguistically, OCP constraints on consonants seem to target adjacent syllables and features with acoustic correlates that are more ‘spread out’ over time. Such features, which include aspiration (e.g. Grassmann's Law) and prenasalisation, are susceptible to perceptual confusion (Ohala Reference Ohala, Masek, Hendrick and Miller1981: 189–196, 1993: 249–257, Gallagher Reference Gallagher2010). That is, it is easy for a listener to perceive such a feature as linked to a segment other than its original host, and in the case of two hosts in adjacent syllables, listeners are likely to attribute the feature's phonetic correlates to a phonological specification on a single segment. The confusion engendered by hosts in adjacent syllables can be resolved by dissimilation (Blevins Reference Blevins2004: 148–149, Bennett & Rose Reference Bennett and Rose2017: 475), or prevented from arising by the OCP. Restrictions against multiple instances of voicing are cross-linguistically rare and phonetically unnatural (Ohala Reference Ohala and Jones1993: 253–254). All known cases of restriction on multiple instances of voicing seem to have arisen diachronically from restrictions on other features (Kawahara Reference Kawahara2008: 327).Footnote 7
We now turn our attention specifically to prenasalisation. On the basis of a wide-ranging survey of the avoidance of /NCVNC/ sequences in several languages, Stanton (Reference Stanton2019) argues that there is a universal dispreference for [NCṼ] sequences, rooted in the perceptual difficulty of assigning nasality to the underlying phonemes in such sequences. If the phonetic details inferred for Old Japanese in §3.2 are essentially correct, prenasalised obstruents in adjacent syllables (/NCV.NCV/) would have produced [NCṼ] sequences ([N.CṼN.CV]). The strong version of Lyman's Law thus fits the cross-linguistic pattern, as long as [NC] can be treated as analogous to [NC]. As Stanton (Reference Stanton2019: 657–658) notes, languages that have salient anticipatory nasalisation on a vowel preceding an onset nasal consonant will also avoid /NCVNV/ sequences, but, as we saw in §3.2, such nasalisation is atypical, and we assume that it did not occur in Old Japanese.
If salient nasalisation had occurred in /NCV.NV/ sequences in Old Japanese, resulting in [N.CṼ.NV], we would expect rendaku to have been avoided in compounds containing E2s such as OJ /tama/ ‘jewel’ and /pana/ ‘flower’, but no such tendency is apparent. In the 62 relevant phonographically attested compound nouns, there are 26 different E2s of the form /CVNV…/. For four of these E2s, the evidence for rendaku is unclear. As noted in §3.4, some phonograms were used inconsistently, sometimes representing a syllable beginning with a ‘voiceless’ obstruent and sometimes representing a syllable beginning with the corresponding prenasalised voiced obstruent. We use the label ‘ambiguous’ to refer to these phonograms. The four E2s in questions are all attested at least once with an E2-initial phonogram that is ambiguous, but not with an E2-initial phonogram that unambiguously represents a prenasalised voiced obstruent. In contrast, twelve of the 26 relevant E2s have at least one token written with an E2-initial phonogram that unambiguously represents a prenasalised voiced obstruent. (See Table I below for a summary.) We can be reasonably confident that these twelve E2s were susceptible to rendaku. The remaining ten E2s are attested only with E2-initial phonograms that unambiguously represent a ‘voiceless’ obstruent. Thus there is no evidence that any of these ten E2s was susceptible to rendaku, although we can never be certain, since they might have had rendaku in unattested compounds.
Table I Old Japanese: distinct /CVNV…/ E2s (N= 26); compounds with /CVNV…/ E2 (N =62); compounds with /CVNC…/ E2 (N=32).
The figures reported in the preceding paragraph and in the following two paragraphs come from an exhaustive search for relevant tokens in the Oxford-NINJAL Corpus of Old Japanese (cf. note 5). The relevant compounds exclude those that would be expected to resist rendaku for some other reason. The excluded examples are: (i) those in which rendaku would violate the strong version of Lyman's (e.g. OJ /suᵑga+para/ ‘sedge plain’); (ii) those with coordinate meaning (e.g. OJ /tuju+simo/ ‘dew and frost’); (iii) those with a numeral E1 (e.g. OJ /ja+kumo/ ‘multilayer [literally ‘eight’] clouds’); (iv) those with honorific OJ /mi/ as E1 (e.g. OJ /mi+tama/ ‘honourable spirit’). The number of phonographically attested tokens of each relevant compound varies widely, ranging from 1 to 51. To identify ambiguous phonograms, we relied on the character values from the table of phonogram usage in Jōdaigo Jiten Henshū Iinkai (1967: 891–903). Whether a phonogram was ambiguous or not can depend on the OJ text in which it occurs.
Of the 62 compounds described above, 17 have at least one token written with an E2-initial phonogram that unambiguously represents a prenasalised voiced obstruent. Some of these 17 compounds almost certainly varied between a form with rendaku and a form without, but there is little doubt that they had or could have rendaku. One of them is OJ /aka+ⁿdama/ ‘red jewel’, used above in §3.4 as example (14c). This compound is attested phonographically twice, once in Kojiki (712), with unambiguous 陀 representing OJ /ⁿda/, and once in Nihon shoki, with unambiguous 娜 representing OJ /ⁿda/. Another 32 of the 62 compounds are attested only with E2-initial phonograms that unambiguously represent a ‘voiceless’ obstruent. One of these is OJ /patu+pana/ ‘first flower’, which is attested phonographically four times in Man'yōshū, with 波 unambiguously representing OJ /pa/ in each token. Of course, some of these compounds might have allowed an alternative form with rendaku, but there is no phonogram evidence to back up such a claim. The remaining 13 compounds are all attested at least once with an E2-initial phonogram that is ambiguous, but not with an E2-initial phonogram that unambiguously represents a prenasalised voiced obstruent. (See again Table I below.) One of these is OJ /pito+tuma/ ~ ?/pito+ⁿduma/ ‘another's spouse’, attested phonographically five times in Man'yōshū, with unambiguous 都 representing OJ /tu/ in two cases and ambiguous 豆 representing OJ /tu/ or OJ /ⁿdu/ in three cases. Philologists agree that most of these 13 had or could have rendaku, and the headword in Jōdaigo Jiten Henshū Iinkai (1967) is given as OJ /pito+ⁿduma/, with rendaku, but the phonogram evidence is equivocal.
In stark contrast to Old Japanese /CVNV…/ E2s, there is virtually no doubt that E2s containing a prenasalised voiced obstruent in the second syllable (OJ /CVNC…/) never had rendaku. In a total of 32 compound nouns, 19 different OJ /CVNC…/ E2s are phonographically attested. For 31 of these compounds, the E2-initial phonogram in every token unambiguously represents a syllable beginning with a voiceless obstruent. (See Table I.) One of these is OJ /aki+kaⁿze/ ‘autumn wind’, attested phonographically four times in Man'yōshū, with unambiguous 可 representing OJ /ka/ in three cases and unambiguous 加 representing OJ /ka/ in one case. The remaining compound combines the E1 OJ /sita/ ‘bottom’ with the E2 OJ /kaze/ ‘wind’, and the sole phonographic attestation of this compound has ambiguous 賀 representing the E2-initial syllable. There is no real doubt that the compound was pronounced OJ /sita+kaⁿze/, without rendaku, but the phonogram evidence is equivocal. The sole phonographic attestation of this compound is in Hitachi fudoki (714–718), and man'yōgana 賀 was ambiguous between OJ /ka/ and /ᵑga/ in this text, as it was in most of the Old Japanese texts in which it was used.
The pattern for /CVNC…/ compounds is clear, and markedly different from those above. We can say with a high degree of confidence that Old Japanese compound nouns with a prenasalised voiced obstruent in the second syllable of E2 never show rendaku. In short, Old Japanese was consistent with Stanton's (Reference Stanton2019) generalisation, assuming, as we have, that an onset nasal did not induce salient nasalisation on an immediately preceding vowel.
5 Northern Tōhoku dialects
5.1 Northern Tōhoku rendaku
Japanese dialects spoken in the northern part of the Tōhoku region of northern Honshū preserve the prenasalisation of voiced obstruents, and it appears that Lyman's Law has remained a ban on prenasalisation in adjacent syllables in some of these dialects. In 2012, fieldwork on rendaku in one representative dialect was carried out in Kahoku-chō, a small town in northern Yamagata Prefecture (Miyashita et al. Reference Miyashita, Irwin, Wilson and Vance2016: 180–183). This survey was led by Miyashita, who is a Kahoku-chō native, and the participants were 24 locally born and raised native speakers (13 male, 11 female), ranging in age from 63 to 94 at the time of recording.
The outcome of rendaku in conservative northern Tōhoku dialects is typically a prenasalised voiced obstruent. The Kahoku-chō survey participants produced many such forms (Miyashita et al. Reference Miyashita, Irwin, Wilson and Vance2016: 183–192), including those in (18).
-
(18)
In most northern Tōhoku dialects, OJ [ᵑg] has shifted to [ŋ], but many older Kahoku-chō speakers retain [ᵑg], as in Kahoku-chō [hɑmᶐᵑgɯɾi] (18b).
Another salient characteristic of northern Tōhoku dialects is that, in most cases, the consonants corresponding to Tokyo ‘standard’ voiceless stops and voiceless affricates are voiced (but not prenasalised) intervocalically in word-medial position, as shown in (19).Footnote 8
-
(19)
The expectation is that a counterpart to Lyman's Law in the Kahoku-chō dialect would involve prenasalised voiced obstruents, and as Kahoku-chō [tẽᵐbɯgɯɾo] in (18a) shows, a simple voiced obstruent in E2 ([g] in this example) does not block rendaku.
5.2 Prenasalisation in adjacent syllables in northern Tōhoku
Modern Tokyo Japanese has the compound /nabe+buta/ ‘pot lid’, which exhibits rendaku (cf. /nabe/ ‘pot’, /ɸuta/ ‘lid’). The corresponding compound is not ordinarily used in the Kahoku-chō dialect. As a result, only 20 of the 24 Kahoku-chō survey participants were able to come up with something like the expected form in response to a picture prompt. (Kahoku-chō speakers do use a phrase corresponding to Tokyo /nabe no ɸuta/, with the two nouns linked by a genitive particle: [nᶐᵐbęnoɸɯtɑ].) Six of these 20 speakers produced a near-standard form with no prenasalisation and no voicing of the stop in the last syllable, as in Table II. Two other speakers had a prenasalised medial stop in E1, but a voiceless medial stop in E2, and another six speakers had a voiced medial stop in E2, but no prenasalisation.
Table II Kahoku-chō speaker responses to picture prompt for E1: [n
mb
] ‘pot’+E2: [
ta] ‘lid’.
The 14 productions in Table IIa all display a lack of full integration into the traditional phonological system of the Kahoku-chō dialect, indicating non-nativeness. Even the oldest of the Kahoku-chō participants were born too late not to be impacted by the relentless standardisation policy of the Japanese national government, which began in the Meiji period (1868–1912). Like most northern Tōhoku speakers, Kahoku-chō speakers are acutely aware that voiced obstruents corresponding to Tokyo voiceless obstruents and prenasalised obstruents corresponding to Tokyo voiced obstruents are salient and highly stigmatised. Most Kahoku-chō speakers, especially those who are educated, have a tacit understanding of the correspondences between the local dialect and the Tokyo standard.Footnote 9 Consequently, they are capable to some degree of converting local forms to standard forms and vice versa, and both the almost fully standardised form and the partially standardised forms in Table IIa are natural outcomes under these circumstances.
The remaining six survey participants who produced a form corresponding to modern Tokyo /nabe+buta/ all had prenasalised [ᵐb] for the medial consonant in E1 and voiced [d] for the medial consonant in E2, but only one had prenasalisation on the initial consonant of E2, as shown in Table IIb. Furthermore, this sole instance of E2-initial prenasalisation is not entirely unambiguous phonetically, as indicated by the question mark.
What is hard to understand about the form Kahoku-chō [nᶐᵐbębɯdɑ] in Table IIb is that it contains both stigmatised features (prenasalisation and stop voicing), but deviates from the correspondence pattern at the rendaku site. The key to understanding the form was provided in a presentation about another northern Tōhoku dialect, spoken in the town of Shizukuishi-chō in Iwate Prefecture (Uwano Reference Uwano2015). In the Shizukuishi-chō dialect, prenasalised obstruents in consecutive syllables are phonotactically prohibited. For example, consider the Shizukuishi-chō compound in (20), which corresponds to Tokyo /hana+bi/ ‘fireworks’ (cf. /hana/ ‘flower’, /hi/ ‘fire’) and shows the expected prenasalisation at the rendaku site.
-
(20)
Now consider the Shizukuishi-chō compound corresponding to Tokyo /kaba+bi/ ‘ceremonial fire’ (cf. /kaba/ ‘birch’, /hi/ ‘fire’), shown in (21). The prenasalisation in the preceding syllable blocks prenasalisation at the rendaku site, although rendaku does occur.
-
(21)
Although the Shizukuishi-chō dialect preserves prenasalised [ᵐb ⁿd ⁿʣ], OJ /ᵑg/ [ᵑg] has become [ŋ], just as in (conservative) Tokyo pronunciation. This [ŋ] does not block prenasalisation in an adjacent syllable, i.e. the constraint has maintained its phonetic grounding. For example, compare two compound verbs corresponding to Tokyo /nige+dasu/ ‘run away’ (22a) and /tobi+dasu/ ‘fly away’ (22b).
-
(22)
In (22b), prenasalisation following E1 would produce the disfavoured phonetic sequence [NCṼ], as in the asterisked form. The constraint against prenasalisation in adjacent syllables prevents this sequence from arising, as explained in §4 in connection with the strong version of Lyman's Law in Old Japanese. In (22a), on the other hand, prenasalisation following E1 does not produce the disfavoured [NCṼ] sequence. We saw in §2 and §4 that, for modern Tokyo speakers who have onset [ŋ], a medial [ŋ] in an E2 blocks rendaku, i.e. [ŋ] behaves like a voiced obstruent with respect to the modern Tokyo version of Lyman's Law. In contrast, onset [ŋ] in the Shizukuishi-chō dialect does not behave like a prenasalised obstruent with respect to the constraint against prenasalisation in adjacent syllables, despite the fact that it derives historically from prenasalised [ᵑg]. Notice that if the same constraint holds in the closely related Kahoku-chō dialect, then [nᶐᵐbębɯdɑ] in Table IIb is actually the fully nativised form corresponding to Tokyo /nabe+buta/.
When we look beyond compounds, however, we find examples that pose a serious challenge to our hypothesis that there is a straightforward ban on prenasalisation in adjacent syllables in the northern Tōhoku dialects we have considered. In the Shizukuishi-chō dialect, the conditional form of a verb ends with a suffix that is ordinarily pronounced /ᵐba/, and some verb stems end in a syllable that has a prenasalised onset consonant. The question that arises, of course, is whether these forms are pronounced with prenasalisation in adjacent syllables. The answer is that these forms are variable, as the examples in (23) show.Footnote 10
-
(23)
We propose that the variability in examples like (23b) can be understood as a conflict between faithfulness (in this case, the desire to avoid allomorphy by maintaining the ‘basic’ form of the suffix) and the constraint prohibiting prenasalisation in adjacent syllables. Since speakers cannot have it both ways, they vacillate as to whether or not to apply dissimilation. In an OT analysis, neither of the two constraints would outrank the other. As the asterisked form in (23b) shows, dissimilation can alter only the initial consonant of the suffix, not the last consonant of the stem, which an OT analysis might handle by ranking root/stem faithfulness higher than affix faithfulness (McCarthy & Prince Reference McCarthy, Prince, Beckman, Dickey and Urbanczyk1995: 364–365).
Although Miyashita is a proficient speaker of the Kahoku-chō dialect (thanks to spending a lot of time with her grandparents as a child), she does not have the secure native intuitions of Kahoku-chō speakers a generation older. As noted in §5.1, many older Kahoku-chō natives retain prenasalised [ᵑg]. Does the Kahoku-chō dialect in fact have a phonotactic constraint prohibiting prenasalisation in consecutive syllables? If so, does Kahoku-chō [ᵑg] block prenasalisation, just like other prenasalised voiced obstruents? And do Kahoku-chō inflectional forms show the kind of variability that we see in Shizukuishi-chō? Unfortunately, these are questions that the 2012 Kahoku-chō survey did not address. Our fragmentary knowledge at the time did not give us any reason to probe in these directions. Since all northern Tōhoku dialects are endangered, we feel extremely fortunate to have learned what we have about the phonotactics of the closely related Shizukuishi-chō dialect, and there is still hope that the necessary follow-up fieldwork in Kahoku-chō can be carried out in the near future.
6 The domain of Lyman's Law in Old Japanese
6.1 Compounds
We have argued in this paper that Lyman's Law in Old Japanese (the ‘strong’ version of Lyman's Law) was a constraint that blocked prenasalised obstruents in adjacent syllables (§3.4), but we have not considered whether there is any evidence that might cast doubt on the claim that adjacency was crucial. In this section, we report a thorough search for examples that could be construed as instances of a non-adjacent prenasalised voiced obstruent inhibiting rendaku. As noted in §3.4, the extant Old Japanese materials have serious limitations, but the results of our search give us no reason to doubt that adjacency was in fact crucial.
As explained in §2, Lyman's Law in modern Tokyo Japanese blocks rendaku when there is a voiced obstruent anywhere in E2. Adjacency is not relevant; a voiced obstruent in E2 blocks rendaku even if it is not in the immediately following syllable. Furthermore, a voiced obstruent in the last syllable of E1 does not block rendaku, even though it is adjacent to the syllable containing the rendaku site.
In northern Tōhoku dialects, on the other hand, the evidence cited in §5.2 from Shizukuishi-chō indicates that what is prohibited is prenasalised voiced obstruents in adjacent syllables. The strong version of Lyman's Law in Old Japanese (§3.4) was also a constraint involving prenasalised voiced obstruents, and the examples cited in §4, repeated in (24), are consistent with the claim that adjacency was relevant.
-
(24)
Each of the examples in (24) contained two prenasalised voiced obstruents, but not in adjacent syllables.
The question that remains to be addressed is whether the strong version of Lyman's Law prevented rendaku only when rendaku would have resulted in prenasalised voiced obstruents in adjacent syllables. The situation in Old Japanese is hard to assess, for two reasons. First, the number of phonographically attested compounds is limited. Second, very few Old Japanese morphemes were longer than two syllables.
One important example is given in (25).
-
(25)
In the sole phonographic attestation of OJ /swoⁿde+tuke+ᵑgoromo/, each of the five obstruent-initial syllables is written with an unambiguous phonogram (see §4). This example shows that a prenasalised voiced obstruent not in the morph immediately preceding the rendaku site did not block rendaku. It is not clear from (25), however, whether the domain of the inhibiting effect in E1 was the entire preceding morph or just the preceding syllable. Distinguishing these two possibilities requires examples with a monomorphemic E1 that contains a prenasalised voiced obstruent somewhere other than in its final syllable. As noted in §3.2, word-initial prenasalised voiced obstruents did not occur in the non-mimetic native Old Japanese vocabulary. Consequently, the compounds of interest must have an E1 that was one of the few Old Japanese morphs longer than two syllables. The sole relevant example is shown in (26).
-
(26)
Rendaku in (26) indicates that a prenasalised voiced obstruent preceding the rendaku site had to be in the immediately preceding syllable to block rendaku. In the sole phonographic attestation of this compound, the phonograms for the two obstruent-initial syllables in E2 are unambiguous (see §4), but the phonogram for the second syllable in E1 is ambiguous, sometimes representing OJ /ta/ and sometimes representing OJ /ⁿda/. The headword in Jōdaigo Jiten Henshū Iinkai (1967) for E1 as an independent word is given as OJ /maⁿdara/, but it is not attested phonographically. Later phonographic attestations (from Middle Japanese) all have a voiced obstruent in the second syllable, matching modern Tokyo /madara/. It is thus very unlikely that example (26) was pronounced /matara+ᵐbusuma/. If it was, it is irrelevant here.
Unfortunately, for E2 there are no convincing diagnostic examples involving a prenasalised voiced obstruent that is not in the syllable immediately following the rendaku site. The absence of rendaku in the examples in (27) is suggestive, but by no means conclusive.
-
(27)
Both E2s in (27) are etymologically composite, although Old Japanese speakers may not have analysed them as such. In the sole phonographic attestation of (27a), E2-initial OJ /po/ and E2-medial OJ /ᵑgi/ are both written with unambiguous phonograms. In the sole phonographic attestation of (27b), the phonogram representing E2-initial OJ /ta/ is unambiguous, but the phonogram representing E2-medial OJ /ᵐba/ is ambiguous, sometimes representing /ᵐba/ and sometimes /pa/. Furthermore, of the 13 phonographic attestations of this E2 as an independent word, this penultimate syllable is written with an ambiguous phonogram in nine cases and with a phonogram that unambiguously represents OJ /pa/ in the other four cases. If the E2 in (27b) was actually pronounced /tatipana/, then the absence of rendaku is irrelevant here.
For OJ /pototoᵑgisu/ ‘cuckoo’ in (27a), Martin (Reference Martin1987: 416) gives the etymology in (28).
-
(28)
Mimetic morphemes are usually characterised as consistently rendaku-immune in modern Tokyo Japanese (Vance Reference Vance2015: 416–417). The immunity may not have been quite so consistent in Old Japanese, but distinction between mimetic and non-mimetic is difficult to draw, because the boundary between these two sectors of the vocabulary has always been fuzzy and porous (Hamano Reference Hamano1998: 6–7). In any case, if the first two syllables of OJ /pototoᵑgisu/ derive from a rendaku-immune morpheme, there is no reason to attribute the absence of rendaku in (27a) to the non-adjacent prenasalised voiced obstruent OJ /ᵑg/.
For OJ /tatiᵐbana/ ‘tangerine’ in (27b), Martin (Reference Martin1987: 543) gives the etymology in (29).
-
(29)
It seems likely that speakers would have seen the morpheme for ‘flower’ in OJ /tatiᵐbana/, and, as long as the entire word was felt to be compound of some sort, its resistance to rendaku could have developed because of the right-branch condition, which restricts rendaku to a right branch in constituent structure (Otsu Reference Otsu1980: 217–222, Vance Reference Vance, Kuno, Makino and Strauss2007b, Kumagai Reference Kumagai2014). The diagrams in (30) illustrate.
-
(30)
In most three-element compounds X + Y + Z, the middle element Y combines with X to form a complex E1, as in (30a), or with Z to form a complex E2, as in (30b). Assuming that neither Y nor Z is immune to rendaku, the right-branch condition predicts that rendaku is possible in both Y and Z in (30a), but only in Z in (30b). If the pre-Old Japanese ancestor of the word for ‘mountain tangerine’ had the structure /jama/ + {/tati+ᵐbana/}, the middle element /tati/ was on a left branch, like Y in (30b). If the right-branch condition was active in pre-Old Japanese, it would have made any complex (Y + Z) E2 immune to rendaku, regardless of whether or not there was a prenasalised voiced obstruent anywhere in that E2. The upshot is that the absence of rendaku in OJ /jama+tatiᵐbana/ (27b) may originally have had nothing to do with the OJ /ᵐb/ in /tatiᵐbana/.
The claim that adjacency was crucial for the strong version of Lyman's Law would be greatly strengthened if there were an example of rendaku in a two-element Old Japanese compound with an E2 that (i) was unambiguously monomorphemic, (ii) had more than two syllables and (iii) contained a prenasalised voiced obstruent in the third syllable or later. Unfortunately, there are simply no such examples. Nevertheless, the available evidence from compounds is consistent with the idea that Lyman's Law in Old Japanese reflected a phonetically motivated prohibition against prenasalised voiced obstruents in adjacent syllables, as proposed in §3.4.
6.2 Inflectional forms and frozen phrases
Just as we saw above in §5.2 in connection with northern Tōhoku dialects, when we look beyond compounds in Old Japanese, there are examples that pose serious problems for our hypothesis that Old Japanese had a straightforward ban on prenasalisation in adjacent syllables. First, there were inflectional suffixes that began with a prenasalised voiced obstruent and attached to verb stems. Since some Old Japanese verb stems ended in a syllable with a prenasalised onset, such inflectional forms were a potential source for prenasalisation in adjacent syllables. Second, the genitive particle OJ /ᵑga/ linked nouns, just like OJ /no/ (see §3.3). Many noun + /ᵑga/ + noun phrases were lexicalised and thus appear as headwords in Jōdaigo Jiten Henshū Iinkai (1967). Since many Old Japanese nouns ended in a syllable with a prenasalised onset, these frozen phrases are another potential source for prenasalisation in adjacent syllables within a phonological word.
The examples in (31) are inflectional forms of verbs that are phonographically attested, and appear to have had prenasalised obstruents in consecutive syllables. A systematic search of the Oxford-NINJAL Corpus of Old Japanese yielded only two other relevant examples. Both are negative forms like (31a) and (b), and thus add little information.
-
(31)
The phonemic transcriptions in (31) obviously violate the proposed ban on prenasalisation in adjacent syllables, but there is room for doubt about whether these transcriptions accurately reflect Old Japanese pronunciation. As noted above in §3.4, some man'yōgana characters were used inconsistently, sometimes representing a syllable beginning with a ‘voiceless’ obstruent and sometimes a syllable beginning with the corresponding prenasalised voiced obstruent. In the single phonographic attestation of (31a), OJ /joᵐba-ⁿzu/, the character that represents the second syllable (婆) was used sometimes for OJ /ᵐba/, but sometimes for OJ /pa/. In the single phonographic attestation of (31b), /iⁿde-ⁿzu/, the character that represents the second syllable (弖) was normally used for OJ /te/, but not for OJ /ⁿde/. There are three phonographic attestations of (31c), /aswoᵐbe-ⁿdomo/, and in two of them the character that represents the fourth syllable (等) was normally used for OJ /to/, but not for OJ /ⁿdo/.
Turning next to Old Japanese frozen noun + /ᵑga/ + noun phrases, (32) shows the only two phonographically attested examples we have found. Like the examples in (31), we found these in a systematic search of the Oxford-NINJAL Corpus of Old Japanese.
-
(32)
The phonemic transcriptions in (32), like those in (31), clearly violate the proposed ban on prenasalisation in adjacent syllables, but here again there is room for doubt whether the transcriptions accurately reflect Old Japanese pronunciation. There is only one phonographic attestation of (32a), OJ /suⁿzu+ᵑga+ne/ (in Man'yōshū), and the phonograms 受 for OJ /ⁿzu/ and 我 for OJ /ᵑga/ are unambiguous. There are four phonographic attestations of (32b), OJ /taⁿdu+ᵑga+ne/. Three of them are from Man'yōshū and have unambiguous 我 (OJ /ᵑga/) for the genitive particle but unambiguous 頭 (OJ /tu/, not OJ /ⁿdu/) for the second syllable. In the fourth phonographic attestation (in Kojiki), the character 豆 for the second syllable was used sometimes for OJ /ⁿdu/ and sometimes for OJ /tu/, and the character 賀 for the genitive particle was used sometimes for OJ /ᵑga/ and sometimes for OJ /ka/.
We suggest that the phonogram spellings described in the preceding two paragraphs can be construed as evidence that dissimilation may have applied to Old Japanese inflectional forms and frozen phrases that would have violated our proposed constraint against prenasalisation in adjacent syllables. We can also surmise that such instances of dissimilation would have been especially susceptible to ‘correction’ by copyists working in later centuries, although we have no evidence to offer. (As noted in §3.4, there are undoubtedly transmission errors in the earliest extant manuscripts of Old Japanese texts, which are handwritten copies of handwritten copies.) It is quite possible, of course, that Old Japanese forms like those in (31) and (32) showed variability of the kind we saw in §5.2 for Shizukuishi-chō inflectional forms like (23b), [jõᵐbɰᵐba] ~ [jõᵐbɯ-ba]). That is, faced with the dilemma of avoiding prenasalisation in adjacent syllables but also avoiding allomorphy, Old Japanese speakers may have wavered, as in (33).
-
(33)
Some of the phonogram spellings noted suggest that, in contrast to the Shizukuishi-chō dialect, even altering a root or stem may have been an option in Old Japanese, as in (34).
-
(34)
We acknowledge that we are speculating here on the basis of very little evidence, but we would rather not simply stipulate that our proposed constraint against prenasalisation in adjacent syllables applied to compounds but not to inflectional forms and frozen phrases in Old Japanese. Inflectional forms like those in (31) and frozen phrases like those in (32) created a conflict between this phonetically motivated constraint and the equally natural desire to avoid allomorphy. The suggestion we have just outlined is that the hypothesised constraint was overridden in such circumstances, but perhaps only variably.
If we compare compounds in which rendaku would have violated the strong version of Lyman's Law, the same conflict would not arise. As we noted in §4, rendaku in modern Tokyo Japanese can be treated as a subsegmental linking morpheme that joins the two elements of a compound. Diachronically, this linking morpheme is the descendant of the prehistoric Japanese genitive particle /no/, as explained in §3.3, and its Old Japanese counterpart was prenasalisation that docked onto the first segment in the second element of a compound, as in (35). The E2-initial syllables in the phonographically attested tokens of examples (35)–(37) are all represented by unambiguous phonograms (in the sense explained above in §4).
-
(35)
The presence or absence of the Old Japanese linking morpheme was unpredictable, just as the presence or absence of its diachronic source (the genitive particle) was unpredictable in prehistoric Japanese (see §3.3). In an example like (36), the linking morpheme was absent.
-
(36)
In the examples in (37), rendaku would have resulted in prenasalisation in adjacent syllables, and we assume that the linking morpheme was predictably absent in such cases.
-
(37)
Just as there is no exponent of the linking morpheme in (36), there is none in (37): (37a) has OJ /k/, not /ᵑg/, and (37b) has OJ /s/, not /ⁿz/. An Old Japanese speaker had no reason to infer underlying forms that contained a morpheme that never surfaced. Rendaku in the examples in (37) would have created both prenasalisation in adjacent syllables and E2 allomorphy; the actual forms had neither. The proposed constraint against prenasalisation in adjacent syllables affected Old Japanese compounds in the sense that it prevented an optional linking morpheme from occurring when its realisation would have produced a violation. Prehistorically, the constraint presumably prevented contraction of genitive Pre-OJ /no/ (see §3.3) from leading to the same undesirable outcome. If such a prehistoric noun + /no/ + noun phrase became conventional, speakers could have either maintained it uncontracted or replaced it with a compound involving the simple juxtaposition of the two nouns.
To summarise our discussion of the proposed prohibition against prenasalised voiced obstruents in adjacent syllables, we saw in §6.1 that compounds are consistent with this statement of the constraint. Specifically, the strong version of Lyman's Law in Old Japanese (§3.4) prevented rendaku in any compound in which a prenasalised voiced obstruent appeared in the syllable immediately preceding or immediately following a potential rendaku site. We have seen in this section, however, that there are inflectional forms of verbs and frozen phrases that presumably were phonological words and yet appear to have contained prenasalised voiced obstruents in adjacent syllables. We cited peculiarities of phonogram usage that cast doubt on whether such sequences were actually permissible, but this evidence is less than compelling. It could be that the strong version of Lyman's Law that we see in Old Japanese reflected a purely phonetic constraint that held in prehistoric Japanese, and that the original constraint had already begun to weaken by Old Japanese times.
7 Conclusion
This paper has looked at three kinds of evidence that support the claim that Lyman's Law originated as a constraint prohibiting prenasalisation in consecutive syllables. First, in §3 and §4, we considered phonetic motivation. Constraints on similar consonants in close proximity generally involve features with phonetic cues that are spread out over time. Prenasalisation is one such feature, and our account of Lyman's Law in Old Japanese is compatible with the results of cross-linguistic research on nasality and perceptual confusion. Second, in §5, we looked at Lyman's Law in some endangered dialects of modern Japanese that, unlike Tokyo ‘standard’ Japanese, still retain prenasalised voiced obstruents. The dialects we examined appear to prohibit these prenasalised consonants from occurring in adjacent syllables. Third, in §6.1, we looked carefully at the compounds recorded in phonograms in Old Japanese texts. None of the relevant examples conflicts with our interpretation of Lyman's Law in Old Japanese as a constraint that applies to adjacent syllables. On the other hand, as we saw in §6.2, some inflectional forms and lexicalised phrases challenge the claim that this constraint held across the board without regard to morphological structure. Nonetheless, the phonogram spellings of the few phonographically attested examples contain enough oddities to cast doubt on whether these forms were truly counterexamples to our hypothesis.
Needless to say, this account of the strong version of Lyman's Law in Old Japanese does not explain how it morphed into the constraint that we find in modern Tokyo Japanese. Modern Lyman's Law applies not to words but to compound elements (morphemes in the case of native elements), and adjacency does not seem to be relevant (§2). Recent studies on the diachronic development of consonant co-occurrence constraints by Coetzee (Reference Coetzee2014) and Gallagher (Reference Gallagher2016) offer some tantalising hints, but we leave this problem for future research.
