Formal syntax and deep history

Supplementary Material
Andrea Ceolin, Cristina Guardiano, Monica Alexandrina-Irimia, Giuseppe Longobardi*

* Correspondence:
Corresponding Author
[email protected]

Keywords: phylogeny, formal syntax, parameters, language reconstruction, biolinguistics

1. Online Repository 1
2. Languages 1
3. The syntactic dataset 3
4. Possible languages 5
5. Heatmap - Syntactic Distances 5
6. PCoAs 7
7. Phylogenetic analysis - UPGMA 11
8. Phylogenetic analysis - Hamming distances 12
9. Phylogenetic analysis – BEAST 2 14
10. Network analysis - NeighborNet 18
11. Phonemic data - the Ruhlen Database 19
12. Ultralocality 20
1. Online Repository
The source code to replicate all the figures and the experiments presented in the paper and in the
Supplementary Material is found in the following online repository (along with other relevant data
and information): https://github.com/AndreaCeolin/FormalSyntax

2. Languages
The 69 languages of the dataset, along with their associated Glottolog
(https://glottolog.org/glottolog/language) and ISO 639-3 codes, the family and subfamily they
belong to, their location and geographic coordinates, are listed in Supplementary Table 1.

Language Label Glottocode Iso 639-3 Top-level family Family Location Latitude Longitude

Afrikaans Afk afri1274 afr IE Germanic Cape Town -33.91 18.42

Archi Arc arch1244 aqc NE Caucasian - Machačkala 42.01 47.26

Barese BA pugl1238 nap IE Romance Bari 41.11 16.87

Basque_Central cB guip1235 eus Basque Guipuzcoan Vitoria-Gasteiz 42.85 -2.68

Basque_Western wB bisc1236 eus Basque Biskayan Bilbao 43.26 -2.93

Bulgarian Blg bulg1262 bul IE Slavic Sofia 42.7 23.32

Buryat Bur buri1258 bua Mongolic E Mongolic Ulan-Ude 51.82 107.61

Calabrese_Northern NCA sout3126 nap IE Romance Verbicaro 39.75 15.19

Calabrese_Southern SCA sout2616 scn IE Romance Reggio Calabria 38.11 15.66

Campano Cam napo1241 nap IE Romance S.M. Capua Vetere 41.08 14.25

Cantonese Can cant1236 yue Sino-Tibetan Sinitic Hong Kong 22.4 114.11

Casalasco CR west2342 egl IE Romance Casalmaggiore 44.98 10.42

Danish Da dani1285 dan IE Germanic Copenhagen 55.68 12.57

Dutch Du dutc1256 nld IE Germanic Amsterdam 52.37 4.89

English E stan1293 eng IE Germanic London 51.51 -0.13

Estonian Est esto1258 ekk Uralic Finno-Ugric Tallinn 59.44 24.75

Even_1 Ev1 even1260 eve Tungusic N Tungusic Kustur 67.79 130.4

Even_2 Ev2 even1260 eve Tungusic N Tungusic Sebyan-Kyuyol 65.29 130.01

Evenki Ek even1259 evn Tungusic NW Tungusic Bomnak 54.71 128.86

Faroese FO faro1244 fao IE Germanic Tórshavn 62.01 -6.77

Finnish Fin finn1318 fin Uralic Finno-Ugric Helsinki 60.17 24.94

French Fr stan1290 fra IE Romance Paris 48.86 2.35

German D stan1295 deu IE Germanic Berlin 52.52 13.4

Greek Grk mode1248 ell IE Greek Athens 37.98 23.73

Greek_Calabria_1 CG1 aspr1238 ell IE Greek Bova 37.99 15.92

Greek_Calabria_2 CG2 aspr1238 ell IE Greek Bova Marina 37.93 15.55

Greek_Cypriot CyG cypr1249 ell IE Greek Larnaca 34.09 33.62

Greek_Salento SaG apul1237 ell IE Greek Calimera 40.24 18.27

Hindi Hi hind1269 hin IE Indo-Aryan New Delhi 28.61 77.21

Hungarian Hu hung1274 hun Uralic Finno-Ugric Budapest 47.5 19.04

1
Icelandic Ice icel1247 isl IE Germanic Reykjavik 64.14 -21.94

Irish Ir iris1253 gle IE Celtic Dublin 53.35 -6.26

Italian It ital1282 ita IE Romance Rome 41.9 12.5

Japanese Jap nucl1643 jpn Japonic - Tokyo 35.69 139.69

Kazakh Kaz kaza1248 kaz Turkic Kipchak Almaty 43.22 76.85

Khanty_1* Kh1 khan1279 kca Uralic Finno-Ugric Kazym 63.7 67.24

Khanty_2* Kh2 khan1279 kca Uralic Finno-Ugric Kazym 63.7 67.24

Korean Kor kore1280 kor Koreanic - Seul 37.57 126.98

Kirghiz Kyr kirg1245 kir Turkic Kipchak Bishkek 42.87 74.57

Lak Lak lakk1252 lbe NE Caucasian - Kumukh 42.54 47.89

Malagasy Mal plat1254 plt Austronesian Malayo-Polinesian Antananarivo 18.88 47.51

Mandarin Man mand1415 cmn Sino-Tibetan Sinitic Beijing 39.9 116.41

Marathi Ma mara1378 mar IE Indo-Aryan Mumbai 19.08 72.88

Mari_1* mM1 mari1278 chm Uralic Mari Shap 56.44 47.96

Mari_2* mM2 mari1278 chm Uralic Mari Shap 56.44 47.96

Norwegian Nor norw1258 nor IE Germanic Oslo 59.91 10.75

Parma PR cent1959 egl IE Romance Parma 44.8 10.32

Pashto Pas pash1269 pus IE Iranian Khyber Pass 34.09 71.16

Polish Po poli1260 pol IE Slavic Warsaw 52.23 21.01

Portuguese Ptg port1283 por IE Romance Lisbon 38.72 -9.1

Reggio_Emilia RE cent1959 egl IE Romance Reggio Emilia 44.7 10.63

Romanian Rm roma1327 ron IE Romance Bucharest 44.43 26.1

Russian Rus russ1263 rus IE Slavic Moscow 55.76 37.62

Salentino Sal pugl1238 scn IE Romance Cellino San Marco 40.47 17.96

Serbo-Croatian SC sout1528 hbs IE Slavic Zagreb 45.82 15.98

Siciliano_Mussomeli MsS cent1963 scn IE Romance Mussomeli 37.57 13.75

Siciliano_Ragusa RGS sout2617 scn IE Romance Ragusa 36.92 14.72

Slovenian Slo slov1268 slv IE Slavic Ljubljana 46.06 14.51

Spanish Sp stan1288 spa IE Romance Madrid 40.42 -3.7

Tamil Ta tami1289 tam Dravidian - Madras 13.08 80.27

Telugu Te telu1262 tel Dravidian - Hyderabad 17.39 78.49

Teramano Ter neap1235 nap IE Romance Teramo 42.66 13.7

Turkish Tur nucl1301 tur Turkic Oghuz Ankara 39.93 32.86

Udmurt_1* Ud1 udmu1245 udm Uralic Permian Chur 57.07 53.03

Udmurt_2* Ud2 udmu1245 udm Uralic Permian Chur 57.07 53.03

Uzbek Uz uzbe1247 uzb Turkic Turkestan Turkic Tashkent 41.3 69.24

Welsh Wel wels1247 cym IE Celtic Cardiff 51.48 -3.18

Yakut Ya yaku1245 sah Turkic N Siberian Turkic Jakutsk 62.04 129.68

Yukaghir Yu yuka1259 yux Yukaghir Kolmic (S Yukaghir) Kolyma 65.5 151.09

Supplementary Table 1. The languages of the dataset. For the languages marked with ‘*’ we encoded two diastratic varieties.

2
3. The syntactic dataset
Supplementary Figure 1 contains the 94 binary nominal parameters used for the experiments
presented in the paper, set in 69 languages spanning across up to 13 traditionally irreducible
Eurasian families.
The table should be read as follows:
1st column: progressive number of the parameters (p1, p2, p3, …)
2nd column: acronym of the parameter
3rd column: name of the parameter
4th column: implicational constraints specifying the conditions for setting the parameter
A detailed list of questions used to determine the state of the parameters and instructions to map
them to Supplementary Figure 1 can be found in:

Crisma, P., C. Guardiano and G. Longobardi (2020), Syntactic parameters and language
learnability. Studi e Saggi Linguistici 58, 99-130. doi: 10.4454/ssl.v58i2.265

3
Supplementary Figure 1. 94 binary nominal parameters set in 69 languages.

4
4. Possible languages
We calculated the number of possible languages generated from the first 30 parameters of Table A
using an algorithm first presented in Bortolussi et al. (2011). This is a breadth-first search algorithm
that keeps track of the number of ‘possible strings’, namely, those strings which do not violate the
implicational rules. The algorithm takes as input two strings of length=1, one with the value “+” and
one with the value “-”.Then, from each string it creates three different strings of length=2, adding
one of the three possible values (“+”, “-” and “0”), so that at the first iteration we have “+/+”, “+/-”,
“+/0” and “-/+”, “-/-”, and “-/0”, for a total of six possible strings. Before the next iteration, only the
strings which are compatible with the implicational structures are kept, while the others are
discarded. The procedure is then repeated, so that at each iteration only the strings which are
compatible are kept.
We limited the analysis to the first 30 parameters because the algorithm has exponential complexity,
and therefore, as every subset of strings needs to be triplicated at each iteration, the algorithm will take
much more time to process every string at each following iteration (see the online repository for the
Python script that we used). Through the algorithm, we calculated that the first 30 parameters used
here generate only 152,448 possible languages (~217) instead of 230. These figures suggest that
calculations of the probability of relatedness based on grammatical structure but neglecting the
pervasive effect of such predictable information could be seriously undermined. We expect the rate of
possible languages to increase at an even lower rate when more parameters are added to the search
space, because they will be potentially constrained by higher numbers of previous parameters.

5. Heatmap - Syntactic Distances

Instructions to visualize the heatmap in Figure 1 in the text.
1. Go to the page https://software.broadinstitute.org/morpheus/
2. Upload to the page the file jaccard_distances.txt from the GitHub repository (link:
https://github.com/AndreaCeolin/FormalSyntax/blob/master/jaccard_distances.txt) and click
the “OK” button to visualize the heatmap
3. In the “Tools” menu, select the option “Hierarchical clustering”, and then the following options:
a. Metric > Matrix values (from a precomputed distance matrix)
b. Linkage method > average
c. Cluster > Rows and columns
Click the “OK” button.
4. To visualize the same color distribution as Figure 1, follow the instructions below:
a. In the “View” menu, select “Options”
b. In the “Color Scheme” window:
i. Uncheck the “Relative color scheme” choice
ii. “Maximum” > 0.857
iii. “Add color stop”
iv. “Selected color” > yellow
v. “Selected value” > 0.430

Supplementary Table 2 lists the maximal clusters of cells in Figure 1 which do not contain any
yellow/red cell. The symbol δ refers to the Jaccard distance between two languages, the symbol μ

5
to the average distance among the languages belonging to a given aggregation/cluster, obtained as
the mean of all the pairwise distances between the languages of that aggregation.
Supplementary Table 3 lists the subgroups which can be identified within each of the cluster in
Supplementary Table 2, along with the distance range and mean within each subfamily.

Maximal cluster δ (range) μ

1. Indo-European From 0 to 0.42 0.26
2. Dravidian+NE Caucasian From 0.10 to 0.25 0.17
3A. Uralic From 0.05 to 0.24 0.16
4A. Altaic+Yukaghir From 0 to 0.32 0.16
3B. Balto-Finnic 0.13 0.13
4B. [rest of]Uralic+Altaic+Yukaghir From 0 to 0.42 0.23
5. Basque 0.17 0.17
6. Sinitic 0.10 0.10
7. Korean-Japanese 0.17 0.17

Supplementary Table 2. Clusters of cells which do not contain any yellow/red cell in Figure 1.

Subfamily δ (range) μ
Romance From 0 to 0.29 0.16
Greek From 0 to 0.17 0.11
Germanic From 0.04 to 0.19 0.12
Slavic From 0 to 0.17 0.08
Indo-Iranian From 0.05 to 0.24 0.16
Dravidian 0.10 0.10
NE Caucasian 0 0
Balto-Finnic 0.13 0.13
Ugric From 0.07 to 0.19 0.14
Permic-Volgaic From 0.05 to 0.11 0.07
Tungusic 0 0
Turkic From 0 to 0.12 0.05

Supplementary Table 3. Distances and means within the subfamilies identifiable in Figure 1.

Other observations:
a. Only one pair formed by a member of Cluster 1 (IE) and a language outside of it has δ<0.26
(i.e. lower than the μ of the cluster), i.e. Ma-Ta (0.25); two pairs have δ=0.26 (Ma-Te and
Hi-Te). Overall, there are 185 white/blue cells (δ < 0.429) involving a member of Cluster 1
and a language outside of it. Most such pairs contain one Indo-Iranian language and one
Dravidian, NE Caucasian, Uralic or Altaic language.
b. All the members of Cluster 2 display many similarities with other languages of the sample:
overall, 93 pairs involving either of the two Dravidian languages and one Indo-European,
Uralic or Altaic language, and 90 pairs involving either of the two NE Caucasian languages
and one Indo-European, Uralic or Altaic language are either white or light blue, with δ
ranging from 0.25 to 0.42.
c. Almost all the languages belonging to Cluster 3A display similarities with many other
languages outside of it (124 blue/white cells), notably Indo-Iranian, Dravidian, NE
Caucasian, Altaic, Yukaghir and (to a smaller extent) Malagasy and Basque, with δ ranging
from 0.19 to 0.42.

6
d. As far as Cluster 4A is concerned, there are 163 blue/white cells involving one of its
members with a language outside of the cluster, and most of them involve Indo-Iranian,
Dravidian, NE Caucasian, Uralic and, marginally, Malagasy and other IE languages.
Buryat and Yukaghir are the outliers of the cluster: yet, no aggregation of blue/white cells
containing either of the two languages displays a μ/δ smaller than those they hold with the
rest of Cluster 4A (see Supplementary Table 4).
e. The languages sharing non-yellow/red cells with Malagasy (an isolate in the Heatmap) are:
Mari_2 (δ=0.39), Udmurt_2 (δ=0.42), Uzbek, Kazakh, Kirghiz, Turkish (δ=0.40).
f. The languages sharing non-yellow/red cells with Cluster 5 (Basque) are: Mari_2
(δ[Basque_Western]=0.41, δ[Basque_Central]=0.40), Marathi (δ[Basque_Western]=0.39)
and Pashto (δ[Basque_Western]=0.41).
g. Sinitic languages (Cluster 6) do not share any white/blue cell with other languages of the
sample, with the exception of Hindi (δ = 0.38).
h. There are two languages which share non-yellow/red cells with Cluster 7, i.e., Greek and
Cypriot Greek (δ = 0.42 with Japanese, and δ = 0.36 with Korean).

Language Cluster (Language) μ/δ

Buryat Cluster 4A (Tungusic, Turkic, Yukaghir) 0.25
Cluster 4B (Ugric, Permic-Volgaic, Altaic, Yukaghir) 0.27
Cluster 2 (Dravidian+NE Caucasian) 0.34
Indo-Aryan [Cluster 1] 0.35
Yukaghir Cluster 4A (Tungusic, Turkic, Buryat) 0.25
Cluster 4B (Ugric, Permic-Volgaic, Altaic) 0.28
Cluster 2 (Dravidian+NE Caucasian) 0.39
Indo-Aryan [Cluster 1] 0.40

Supplementary Table 4. Relations between Buryat/Yukaghir and closest languages.

6. PCoAs
The PCoAs have been produced using the software PAST
(https://www.nhm.uio.no/english/research/infrastructure/past/). After the distance matrix is
loaded, the following option should be selected: Multivariate -> Ordination -> PCoA.
In the scatter plot, the attribute Row Labels must be selected to display the name of the languages.
The PCoA in Supplementary Figure 2 was obtained from the parametric Jaccard distances
between the 30 non-Indo-European languages of our sample.
In Supplementary Figure 2, the first coordinate, which accounts for about 58% of the variance,
separates Uralic, Dravidian, NE Caucasian, Altaic, and Yukaghir (left area) from the others.
a. Left area: the second coordinate (accounting for 17% of the variance) separates Altaic (with
Buryat falling precisely on the horizontal axis) and Yukaghir (bottom quadrant) from the rest.
In the top quadrant, Uralic, Dravidian and NE Caucasian are not clearly separated: this reflects
the high amount of similarities among these languages observed in the Heatmap.
b. Right quadrant: the second coordinate separates the languages of the Far-East (bottom quadrant)
from the rest. Japanese and Korean, which appear very close to one another in the Heatmap, in
this representation are quite separated.

7
As it appears in the graph, distances, especially in the left quadrant, are quite compressed: hence, the
internal distribution of the pairs does not emerge clearly. In order to observe it in more detail, we
visualized the two groups identified by the first coordinate as two separate graphs, shown in
Supplementary Figure 3 and Supplementary Figure 4.
The distribution of the pairs in Supplementary Figure 3 further emphasizes the neat separation
between Sinitic and Japanese-Korean.
In Supplementary Figure 4:
a. Dravidian and NE Caucasian are a separate cloud (top right quadrant).
b. The top left quadrant shows two major clouds:
i. all Altaic languages but Buryat
ii. Buryat (that expectedly appears as an outlier of the group) and Yukaghir (that, again, is
attracted by the Altaic group)
c. Uralic forms a relatively compact cloud in the bottom area of the graph, with Estonian and
Finnish in an outlying position, as seen in the Heatmap.

Finally, Supplementary Figure 5 contains the 39 IE languages of our sample. Their distribution
partitions the known subfamilies with a discrete resolution and without historical errors. The first
coordinate, which accounts for 46% of the variance, separates Romance from the other
subfamilies. In the left area, the horizontal axis (which accounts for 18% of the variance) identifies:
a. Germanic and Slavic, which form two separate clouds in the bottom-left quadrant
b. Celtic, Greek and Indo-Iranian (more scattered)

Supplementary Figure 2. PCoA of the 30 non-Indo-European languages.

8
Supplementary Figure 3. PCoA of 7 non-Indo-European languages.

9
Supplementary Figure 4 - PCoA of 18 non-Indo-European languages.

10
 Supplementary Figure 5 - PCoA of the 39 Indo-European languages.

7. Phylogenetic analysis - UPGMA

The UPGMA tree (Figure 3 in the main text) has been generated using a modified bootstrapping
procedure.
The bootstrapping technique resamples the whole dataset by selecting each character with equal
probability and recreating a matrix of the same length. The content of the new matrix is different
from the original matrix, because some characters might be absent and some others might be
present multiple times as a consequence of the sampling procedure. This allows one to estimate
the robustness of the dataset by repeating the same analysis on different samples of the dataset.
Since the Jaccard distance between two languages excludes all parameters that are set to '0' in
either one of them, a standard bootstrapping procedure runs the risk of making a pair of languages
not comparable, because in some replicas the number of identities plus differences can reduce to
zero, and then yield a zero denominator for the Jaccard formula. For this reason, we decided to
adopt a moderated bootstrap procedure, by creating 1000 datasets in which only six parameters are
resampled. Since the minimum number of comparable parameters between any two languages in
the dataset is seven, a resampling of six parameters will assure that the two strings are comparable
by means of the Jaccard distance.
The UPGMA tree presented in the text is a consensus tree resulting from applying UPGMA to the
1000 replicas of the dataset.

11
The first two splits of Figure 3 identify the following nodes:
a. The languages spoken in East Asia, with Japanese and Korean falling under one and the same node
b. Basque
A further split separates two major clusters, internally articulated as follows:
1. a. Malagasy
b. Uralic, articulated into the following groups:
- Balto-Finnic
- Ugric
- Volgaic-Permic, with a low bootstrapping score, which shows that the two subfamilies
are often mixed when replicating the experiment
2. a. Altaic+Yukaghir, with the following internal articulation
- Yukaghir is the outlier
- Buryat
- Tungusic
- Turkic: Kazakh and Kirghiz are clustered together, followed in succession by Turkish,
Yakut and Uzbek (NE Turkic). Note the low bootstrapping score of the Kazakh and
Kirghiz node, which means that by replicating the experiment they might end up
clustering with Turkish first.
b. i. NE-Caucasian and Dravidian
ii. Indo-European, articulated into the following major subfamilies:
- Indo-Iranian. Pashto is the outlier. The two Indo-Aryan languages are together
- Romance. Romanian is the outlier. The Ibero-Romance unit (Spanish and Portuguese) is
recognized. The dialects of Italy, and Italian, are under the same node, with the following
internal articulation: Northern Gallo-Italic dialects (Casalasco, Parma and Reggio_Emilia);
Extreme-southern dialects (Salentino, Calabrese_Southern and Siciliano); Upper-southern
dialects (Teramano, Barese, Campano and Calabrese_Northern) and Italian
- Celtic
- Greek. Greek clusters with Greek_Cypriot; Greek_Calabria_1 is the outlier of this
group, reflecting its documented conservative nature (Guardiano et al. 2016,
Guardiano and Stavrou 2014, 2019)
- Slavic. Bulgarian occurs as the outlier. Polish and Russian fall together
- Germanic. Three out of four traditional West-Germanic languages (Continental West
Germanic) are under one and the same node (Afrikaans, Dutch and German). English
falls within the North-Germanic cluster (Icelandic, Danish, Faroese, Norwegian)

8. Phylogenetic analysis - Hamming distances

We created a UPGMA tree (Supplementary Figure 6) from a matrix of Hamming distances, using
the same procedure as for Figure 3. The tree retrieves most of the nodes observed in Figure 3,
with three major differences:
a. Pashto and Romanian go together
b. the nodes containing the two Basque varieties, the two Sinitic languages and Japanese/Korean
are not the outliers (they are closer to the Indo-European node)
c. West-Germanic and a North-Germanic node are identified

12
Supplementary Figure 6. UPGMA tree calculated using Hamming distances.

13
9. Phylogenetic analysis – BEAST 2
In order to determine the best model for the BEAST tree (Bouckaert et al. 2019), we used the
software Tracer (https://beast.community/tracer) to compare the posterior likelihood of several
models. The analysis is summarized in Supplementary Figure 7. The best model that we
determined is a Gamma Site Model with Substitution Rate = 1, a Mutation Death Model with death
p = 0.1, a Relaxed Clock (Logarithmic) with clock rate = 1, and a uniform Yule model for the birth
rate. The Monte Carlo Markov Chain produced 10,000,000 trees, 25% of which were used for the
burn-in and discarded for the purpose of the calculation of the consensus tree. The tree is a
consensus tree of 7,500 different trees sampled through the 7,500,000 trees (with a sample stored
every 1000 generated trees) produced by Monte Carlo sampling.

Supplementary Figure 7. Tracer analysis for different BEAST models used to generate a tree from the syntactic
dataset.

The constrained BEAST tree (Figure 4 in the text) identifies the following splits:
a. The languages spoken in East Asia, with Japanese and Korean falling under one and the same
node
b. Malagasy and Basque
c. Uralic, articulated into the following groups:
- Balto-Finnic (Estonian and Finnish)

14
 - Ugric (Hungarian and Khanty)
- Volgaic-Permic (Mari and Udmurt), with a low posterior probability, which means that the
two subfamilies can appear mixed in some replications of the experiments
d. A node that splits into the following:
- Dravidian (Tamil and Telugu) + NE Caucasian (Archi and Lak)
- Altaic+Yukaghir, with the following internal articulation:
- Buryat
- Yukaghir
- Tungusic (Even, Evenki)
- Turkic: Kazakh and Kirghiz are clustered together, followed in succession by Turkish,
Yakut and Uzbek. All these nodes have low posterior probability, which means that the
internal articulation of the family is not defined, and therefore is not stable across different
replications
e. Indo-European, articulated into the following major subfamilies:
- Indo-Iranian. Pashto is the outlier of the two Indo-Aryan languages (Hindi, Marathi)
- Romance. Romanian is the outlier. French is the outlier of a node that also includes the
Northern Gallo-Italic dialects (Casalasco, Parma and Reggio_Emilia). Salentino is the
outlier of a node that has the following splits: Italian and Upper southern dialects of Italy
(Calabrese_Northern, Teramano, Barese, Campano); Ibero-Romance and the Extreme
southern dialects of Italy (Calabrese_Southern, Siciliano_Mussomeli, Siciliano_Ragusa),
with the exception of Salentino)
- Celtic (Irish and Welsh) + Greek (with the same subarticulation as in UPGMA)
- Slavic (with the same subarticulation as in UPGMA)
- Germanic, split into West- vs. North-Germanic (contrary to UPGMA, both nodes are
correctly identified)

Supplementary Figure 8 displays an unconstrained tree generated using BEAST. Here, Finnish and
Estonian do not cluster with the other Uralic languages, but are the outliers of a group containing
Uralic, NE Caucasian and Dravidian, Turkic, Tungusic, Buryat and Yukaghir. In other replications,
Balto-Finnic appears as an outlier of the Indo-European languages, or even inside this family.
A tree without Finnish and Estonian is displayed in Supplementary Figure 9.

15
Supplementary Figure 8. Unconstrained BEAST tree.

16
Supplementary Figure 9. BEAST tree without Finnish and Estonian.

17
10. Network analysis - NeighborNet
For the network analysis, we used the software SplitsTree (Huson and Bryant 2006) and the
algorithm NeighborNet. The network (Supplementary Figure 10) identifies all the major
aggregations already identified in the other experiments. The two graphs containing the Δ-scores
(Supplementary Figure 11) and the Q-residuals (Supplementary Figure 12) have been produced
using matplotlib in Python3. Supplementary Table 5 lists the ten highest Δ-scores and Q-
residuals for our dataset.

Δ-scores Q-residuals
Mandarin 0.387 Mandarin 0.125
Cantonese 0.387 Cantonese 0.125
Korean 0.371 Japanese 0.107
Japanese 0.369 Korean 0.098
Pashto 0.367 Hungarian 0.097
Basque_Central 0.365 Lak 0.092
Tamil 0.350 Archi 0.092
Basque_Western 0.349 Basque_Central 0.089
Hungarian 0.336 Basque_Western 0.085
Malagasy 0.336 Tamil 0.081

Supplementary Table 5. The ten highest Δ-scores and Q-residuals for our dataset.

Supplementary Figure 10. NeighborNet network obtained using SplitsTree on the syntactic dataset.

18
 Supplementary Figure 11. Δ-scores derived from the network in Supplementary Figure 10.

Supplementary Figure 12. Q-residuals derived from the network in Supplementary Figure 10.

11. Phonemic data - the Ruhlen Database

The Tracer analysis for the tree generated from Ruhlen’s dataset is summarized in Supplementary
Figure 13. The best model that we determined is a Gamma Site Model with Substitution Rate =
1, a Mutation Death Model with death p = 0.1, a Relaxed Clock (Logarithmic) with clock rate = 1
with clock rate = 1, and a uniform Yule model for the birth rate. The Monte Carlo Markov Chain
produced 10,000,000 trees, 25% of which were used for the burn-in and discarded for the purpose
of the calculation of the consensus tree. The tree is a consensus tree of 7,500 different trees sampled
through the 7,500,000 trees (with a sample stored every 1000 generated trees) produced by Monte
Carlo sampling.

19
 Supplementary Figure 13. Tracer analysis for different BEAST models used to generate a tree from the subset of
the Ruhlen dataset overlapping with our languages.

12. Ultralocality
The Network of Supplementary Figure 14 has been generated from the Romance languages of
the sample. Here, the languages of Italy are separated from the rest of Romance, and their internal
classification is largely the expected one: the Lausberg dialect (Calabrese_Northern) is an isolate
bridging the other Upper southern dialects (Campano, Barese and Teramano) and Italian; the
Northern Gallo-Italic group (Teggio_Emilia, Parma and Casalasco) is singled out; the Extreme
southern dialects (Siciliano and Calabrese_Southern) are together, with Salentino as the outlier;
the position of the Extreme southern group suggests some relation with Ibero-Romance.
In the Heatmap in Supplementary Figure 15, white and blue cells mark distances ranging from
0 to 0.142, yellow and red cells mark distances ranging from 0.143 to 0.286.
Instructions to visualize the heatmap:
1. Go to the following page: https://software.broadinstitute.org/morpheus/
2. Upload to the page the file jaccard_distances.txt from the GitHub repository (link:
https://github.com/AndreaCeolin/FormalSyntax/blob/master/Romance/jaccard_distances_rom
ance.txt ), and click the “OK” button to visualize the heatmap.
3. In the “Tools” menu, select the option “Hierarchical clustering”, and then the following:
a. Metric > Matrix values (from a precomputed distance matrix)
b. Linkage method > average

20
 c. Cluster > Rows and columns
Click the “OK” button.
4. To visualize the same color distribution as Fig.1, follow the instructions below:
a. In the “View” menu, select “Options”
b. In the “Color Scheme” window:
i. Uncheck the “Relative color scheme” choice
ii. “Maximum” > 0.286
iii. “Add color stop”
iv. “Selected color” > yellow
v. “Selected value” > 0.143

Supplementary Table 6 lists the maximal clusters of cells in Supplementary Figure 15 which
do not contain any yellow/red cell. The symbol δ refers to the Jaccard distance between two
languages, the symbol μ to the average distance among the languages belonging to a given
aggregation/cluster, obtained as the mean of all the pairwise distances between the languages of
that aggregation.

Maximal cluster δ (range) μ/δ

1. Extreme southern dialects of Italy From 0 to 0.13 0.06
2. Upper southern dialects of Italy and Italian From 0 to 0.12 0.07
3. Northern Gallo-Italic dialects From 0 to 0.08 0.05
4. Ibero-Romance 0.04

Supplementary Table 6. Clusters suggested by the distribution of the distances in the Heatmap in Supplementary
Figure 15.

The white/blue cells outside of the clusters in Supplementary Table 6 correspond to the pairs
listed in Supplementary Table 7.

Cluster - Language Closer Cluster - Language (δ)

1 - Siciliano_Ragusa 2 - Italian (0.08), Calabrese_Northern (0.12), Barese (0.12), Campano (0.12)
4 - Portuguese (0.12)
1 - Siciliano_Mussomeli 2 - Italian (0.08), Calabrese_Northern (0.12), Barese (0.12), Campano (0.12)
4 - Portuguese (0.12)
1 - Calabrese_Southern 2 - Italian (0.08)
4 - Portuguese (0.08)
1 - Salentino 2 - Barese (0.13), Campano (0.13), Calabrese_Northern (0.13), Italian (0.13)
2 - Barese 1 - Siciliano_Ragusa (0.12), Siciliano_Mussomeli (0.12), Salentino (0.13)
3 - Reggio_Emilia (0.13), Parma (0.13)
2 - Campano 1 - Siciliano_Ragusa (0.12), Siciliano_Mussomeli (0.12), Salentino (0.13)
3 - Reggio_Emilia (0.13), Parma (0.13)
2 - Teramano 3 - Reggio_Emilia (0.08), Parma (0.08)
2 - Calabrese_Northern 1 - Siciliano_Ragusa (0.12), Siciliano_Mussomeli (0.12), Salentino (0.13)
2 - Italian 1 - Siciliano_Ragusa (0.08), Siciliano_Mussomeli (0.08), Calabrese_Southern (0.08),
Salentino (0.13)
3 - Reggio_Emilia (0.12), Parma (0.12)
4 - Spanish (0.14)
Romanian (0.14)
3 - Reggio_Emilia 2 - Barese (0.13), Campano (0.13), Teramano (0.08), Italian (0.12)

21
 3 - Parma 2 - Barese (0.13), Campano (0.13), Teramano (0.08), Italian (0.12)
4 - Portuguese 1 - Siciliano_Ragusa (0.12), Siciliano_Mussomeli (0.12), Calabrese_Southern (0.12)
2 - Italian (0.11)
4 - Spanish 2 - Italian (0.14)
(isolate) Romanian 2 - Italian (0.14)

Supplementary Table 7. White/blue cells in Supplementary Figure 15 outside of the clusters listed in
Supplementary Table 6.

In the PCoA in Supplementary Figure 16, the vertical axis (43% of the variance) separates Ibero-
Romance and the dialects of central/southern Italy from the rest of Romance, with the exception of
Teramano; Barese and Campano fall precisely on the vertical axis. The horizontal axis (36% of the
variance) separates the dialects of Italy from the other Romance languages, with two exceptions:
Calabrese_Southern (that appears right below the axis), and Casalasco (the Northern Gallo-Italic
dialect closest to French).

Supplementary Figure 14. Network of the Romance languages.

22
Supplementary Figure 15. Heatmap of the Romance languages.

23
Supplementary Figure 16. PCoA of the Romance languages.

24