给语言认亲戚

像不像不算数，对得上才算数

先把"像"这条线索废掉。除了波斯语那个 bad，课堂上还必讲另一对：希腊语的 theos 和拉丁语的 deus 都表示"神"，长得像、意思同，却不是亲戚，两个词各有各的来源。拉丁语 deus 真正的亲戚是希腊语里的 Zeus（宙斯）——那个一点也不像的。长得像的不一定是一家，一家的不一定长得像。

算数的是成批的、能重复的对应。拉丁语词首的 f，在西班牙语里变成了一个写出来但不发音的 h：filius 对 hijo（儿子）、facere 对 hacer（做）、ferrum 对 hierro（铁）。不是这三个词凑巧，是这个位置上的 f 整批走了同一条路，而意大利语和法语没走（figlio、fils）。判据因此不是"两个词像不像"，是"同一个音在同一个位置上，是不是稳定对着另一个音"。

更漂亮的是连例外都守规矩。拉丁语 focus（炉火）到西班牙语是 fuego，f 好端端留着；可同一个词根派生出的 focaria 到西班牙语是 hoguera（篝火），f 变成了 h。差别在后面跟着什么：f 后面接 ue 这个双元音时不变，别的位置才变。看起来被打破的规律，往往只是条件还没写全。

这条判据还顺手把借来的词踢了出去。西班牙语的 fiesta（节日）、febrero（二月）开头的 f 都没变，因为它们不是一路传下来的，是后来从书面拉丁语借回去的，错过了那场变化。英语里更显眼：本土词是 father、fish、foot，从拉丁语一侧借来的同源词是 paternal、piscine、pedal，同一个位置一边 f 一边 p。一个词开头的辅音就能告诉你它从哪扇门进来——对应关系不只是证据，还是筛子。

EN Similarity is the first thing to throw out. Persian bad means "bad" and sounds like English bad, yet the two are unrelated: the Persian word goes back to an earlier wat. Greek theos and Latin deus both mean "god" and look alike, but they descend from different roots; the real relative of deus is Greek Zeus, which looks nothing like it. What counts instead is a repeatable correspondence across many words. Latin initial f shows up in Spanish as a written but silent h: filius/hijo, facere/hacer, ferrum/hierro, while Italian and French keep the f (figlio, fils). Even the exceptions obey conditions: Latin focus gives Spanish fuego with f intact, but the derived focaria gives hoguera with h, because f stayed put before the diphthong ue. An apparent exception usually just means the conditioning has not been written out yet. The same test throws out loans. Spanish fiesta and febrero keep their f because they were re-borrowed from written Latin after the change had run. English shows it more starkly: native father, fish, foot against borrowed paternal, piscine, pedal — f on one side, p on the other. The first consonant of a word tells you which door it came in through.

拿几种后代语言，推一个没人听见过的词

手艺本身不神秘。先把同源词（cognate，来自同一个祖先形式的一组词）成组排好：几种语言里"儿子"各怎么说，"做"各怎么说。再把每组对得上的音抽出来，看它在别的组里是不是照样对。最后挑一个祖先形式：从它出发，用最少、最自然的几步能走到所有后代。方向也有讲究——夹在元音之间的塞音容易变软，反过来很少见。

凭什么信这套推理？因为有一个家族可以对答案：罗曼语这一支的祖先留下了大量文字。你可以假装不知道拉丁语，只拿西班牙语、意大利语、法语、葡萄牙语、罗马尼亚语反推一遍，再翻书核对。结果对得上，但有个很说明问题的偏差：推出来的不是教科书上那个古典拉丁语，更接近当时人嘴里说的那一版。比较法给你的永远是"这些后代的最近共同祖先"，不是同时代的书面标准。

这套手艺也不是只能用在欧洲。中古汉语的读音同样是被重建出来的：601 年的《切韵》用两个字拼出第三个字的音，把字归了类，却没说那些类到底念什么。高本汉（Bernhard Karlgren）把三份证据对起来——《切韵》的分类、各地现代方言的读法、日语韩语越南语当年整批借走的汉字读音——才把音值填进那张表。方案至今仍在修订，做法是同一套。

最硬的一次检验发生在 1879 年。索绪尔（Ferdinand de Saussure）处理印欧语里元音长短的交替时提出：得假设有一类当时任何已知语言里都找不到的音，这些交替才解释得通。这个主张在半个世纪里被当成外行人的怪念头。二十世纪初，赫梯语（Hittite，小亚细亚出土的一门古代语言）的泥板被解读出来；1927 年 Kuryłowicz 指出，赫梯语里有个辅音正好出现在索绪尔预言的那些位置上。重建不是把旧材料换种写法，它是一个承诺：哪天挖出一门新的亲戚语言，它必须长成这样。

EN The craft itself is plain. Line up cognates — words descended from one ancestral form — in sets: how each language says "son", how each says "do". Pull out the sounds that line up, then check whether they line up the same way in other sets. Finally pick the ancestral form from which the fewest and most natural steps reach every descendant; direction matters too, since stops between vowels tend to soften rather than harden. Why trust this? Because one family lets you check the answer. Reconstruct from Spanish, Italian, French, Portuguese and Romanian, then open a Latin grammar. It works — with one telling offset: what comes out is not the classical literary language but something closer to what people actually spoke. The method always yields the latest common ancestor of the daughters, not the written standard of the same era. Nor is it a European tool. Middle Chinese pronunciation was likewise reconstructed: the Qieyun (切韻, a rime dictionary of 601) sorted characters into classes without saying how the classes sounded, and Bernhard Karlgren filled in sound values by triangulating three sources — those classes, modern regional pronunciations, and the Chinese readings borrowed wholesale into Japanese, Korean and Vietnamese. The hardest test came in 1879, when Ferdinand de Saussure argued that a class of sounds found in no known language had to be posited for Indo-European vowel alternations to make sense. For half a century this was treated as an eccentricity. Then Hittite tablets were deciphered, and in 1927 Kuryłowicz pointed out that a Hittite consonant turns up exactly where Saussure had predicted. A reconstruction is not a restatement; it is a commitment about what a newly unearthed relative would have to look like.

往回推大约八千年就推不动了

既然这么有力，为什么没人把全世界的语言接成一棵大树？因为信号会耗光。每过一千年，一批同源词被别的词替换掉，音变再叠几层，对应越来越难辨认；与此同时，碰巧相似的词一直按固定概率存在。到某个深度，真信号降到噪声以下——不是找不到相似，是相似不再能说明任何事。

噪声有多大是可以算的。Ringe 1992 年做的就是这件事：按纯概率估算两门语言的词表之间能撞出多少对"音近义近"，结果比多数人的直觉高得多。他针对的是把上百种语言的词表摆在一起、靠眼睛找相似的做法。争论至今没完全平息，但有一头是清楚的：没有系统对应撑着的相似，数量上和偶然区别不开。

于是天花板大致落在八千年上下——常被引的说法是 8000 加减 2000 年，上限不超过一万年，依据之一是已知最老的那些语系大概就是这个岁数。再往前的方案，比如把印欧语、乌拉尔语接成更大的"诺斯特拉语系"，处境不是被证伪，是证据强度够不着：那个深度上，能支持它的和偶然产生的长得一样。

今天的前沿是把生物学那套统计建树的工具搬过来，拿同源词的有无当性状，跑出带年代的树。印欧语的起源因此吵了二十多年：一派把根扎在安纳托利亚、跟着农业扩散走，一派扎在东欧草原、跟着古人骨 DNA 的迁徙证据走。2023 年 Heggarty 等人在 Science 上给出约 8120 年的根年代，提了个混合方案，两边各占一段。证据强度得标清楚：这是模型给的后验区间，换一套同源词判定或速率假设，年代会跟着动。

动多少？看汉藏语系。2019 年有两组人用同类方法做同一个语系，地理结论接近，年代一组约 7200 年、另一组约 5900 年，差了一千多年。方法是真的，误差棒也是真的。天花板同理：不是手艺不够，是信息被时间吃掉了，再精巧的算法也读不出不存在的信号。

EN Why has nobody joined all the world's languages into one tree? Because the signal runs out. Every millennium, some cognates are replaced and further sound changes pile up, making correspondences harder to see, while chance look-alikes keep occurring at a fixed rate. Past some depth the real signal drops below the noise: the problem is not that resemblances are missing, but that they no longer show anything. The size of that noise can be computed, and Ringe did exactly that in 1992, estimating how many sound-and-meaning matches two wordlists should throw up by chance alone — far more than most people's intuition. His target was mass comparison, the practice of laying out wordlists from a hundred languages and looking for similarities by eye. The dispute has never fully settled, but one side of it is clear: resemblance unsupported by systematic correspondence cannot be told apart from accident. The ceiling therefore sits around eight thousand years — often quoted as 8000 plus or minus 2000, with an upper bound near ten thousand, partly because the oldest securely established families are about that old. Deeper proposals such as Nostratic or a reconstructed single ancestor for all languages are not refuted; they simply cannot reach the evidential bar. The current frontier borrows statistical tree-building from biology, coding cognates as characters and producing dated trees. Indo-European origins have been argued over for two decades on this basis: an Anatolian homeland tracking the spread of farming against a steppe homeland tracking ancient DNA. In 2023 Heggarty and colleagues published a root age of about 8120 years in Science and proposed a hybrid of the two. Strength: that is a model posterior, and it moves when cognate judgements or rate assumptions change. How much? In 2019 two teams used comparable methods on Sino-Tibetan and landed on about 7200 and about 5900 years, with similar geography and over a millennium between the dates. The method is real and so are the error bars. So is the ceiling — not a failure of craft but information eaten by time.