Source-linked AI summary
Is Luke the Author of a Gospel and the Acts of the Apostles?
Jacques Savoy
TL;DR
Whether Luke’s Gospel and Acts share an author remains disputed, with earlier studies relying largely on qualitative judgments or selected word-frequency differences. The study applies stylometric attribution and verification models, and both indicate common authorship across the tested configurations.
Problem
Previous studies of whether Luke’s Gospel and Acts share an author have relied largely on qualitative judgments or selected word-frequency differences, and a definitive conclusion has not been reached.
Source: §1 · p. 2, §4 · p. 7, §8 · p. 18Method
The study applies Burrows’ Delta at eleven feature-set sizes and an authorship verification model across stylistic representations, feature sizes, and similarity measures.
Source: §8 · p. 18Results
Both Burrows’ Delta and the verification model indicate common authorship of Luke’s Gospel and Acts across the tested configurations.
Source: §8 · p. 18Takeaways & Limitations
The experiments indicate that reliable assessment should use at least 100 stylistic features and vary underlying parameters rather than rely on a few words.
Source: §8 · p. 18Synoptic literary dependence and copyist-introduced textual variation can interfere with the stylistic signal and make identifying the genuine writer harder.
Source: §8 · p. 18
Abstract
from arXiv · showhide
According to Christian tradition, Luke is credited with authoring a Gospel and the Acts of the Apostles, even if his name does not appear in either book, both originally written in Koine Greek. Several biblical scholars assume that both texts were written by a common author, while others deduce the presence of two authors. Different studies have been found to support either finding, some based on qualitative evaluation, while a few others consider the occurrence frequency differences between the two books. To propose an enhanced quantitative analysis, this study is grounded on two recent authorship attribution models. The Burrows' Delta, applied with eleven different feature sizes, demonstrates common authorship. An author verification model confirms this finding. The following experiments consider several stylistic representations, feature sizes, and distance functions to confirm that Luke is the true author of both books.
1. Introduction
Christian tradition credits Luke with both the Gospel and Acts, but scholars have disputed whether one author wrote both. This study tests common authorship with recent stylometric models without assuming that the author was Luke.
1. Introduction
- The shared dedication to Theophilus, narrative continuity, and common themes have been cited as arguments for common authorship.The books also share emphases on the Holy Spirit, sermons for different audiences, communities outside Judea, and Christian innocence.Source: §1 · pp. 1–2
- Common authorship does not by itself establish that Luke and Acts were written together or should be treated as one work.The paper also notes open questions about their narrative, canonical, genre, and theological unity.Source: §1 · p. 2
- The Gospel and Acts are commonly attributed to Luke, but scholars have questioned common authorship, citing differences in word frequencies and style.Arguments for unity include shared dedication, narrative continuity, and thematic relationships between the books.Source: §1 · pp. 1–2
- The study evaluates whether Luke and Acts share an author using recent stylometric models, while not assuming that the author’s name was Luke.The authors state that they largely suppose a single common author wrote both books.Source: §1 · p. 3
2. State of the Art
The paper reviews several stylometric approaches, from frequent-word comparisons to multivariate and machine-learning methods. It also distinguishes authorship attribution from verification and surveys alternative stylistic representations.
2. State of the Art
- Burrows’ Delta compares the most frequent word-tokens using Manhattan distance on Z scores; other methods use predefined functional-word sets and Kullback-Leibler divergence.The review also describes Labbé’s method, which uses the entire vocabulary with a variant of Manhattan distance.Source: §2 · p. 3
- Multivariate approaches project text representations into a reduced space, where texts by the same author are expected to cluster together.Examples include principal component analysis, hierarchical clustering, and discriminant analysis, often using frequent words and POS information.Source: §2 · p. 3
- Machine-learning approaches include decision trees, k-nearest neighbours, neural networks, random forests, logistic regression, and support vector machines.Which performs best depends on the test collection and representation strategy.Source: §2 · p. 3
- Stylometric representations can use character n-grams or POS distributions, which may provide useful complementary information but can be harder to interpret or less performant.Character-sequence methods can perform well, while POS-based approaches assume an effective tagger.Source: §2 · p. 4
- Authorship verification tests whether a disputed text was written by a candidate author, using chunks of the query text and, in some variants, impostor authors.The reviewed approaches aggregate chunk classifications or similarities, sometimes across repeated random feature subsets.Source: §2 · p. 4
3. Evaluation Corpus
The evaluation corpus includes New Testament texts in Greek and English, with the original Luke and Acts in Koine Greek. The paper accounts for synoptic overlap and textual variation that may complicate stylistic comparison.
3. Evaluation Corpus
- The corpus comprises 190 chapters downloaded from Bibelwissenschaft.de, with Greek texts from the Nestle-Aland edition and English examples from the King James version.The texts were preprocessed to remove elements such as verse numbers and running titles.Source: §3 · p. 4
- POS tagging can be ambiguous: the word “still,” for example, may be an adverb, noun, adjective, or verb.The study defines “Greek” as Koine or Ancient Greek for simplified notation.Source: §3 · p. 4
- The selected books include texts with more than 10,000 tokens, while reliable attribution is described as difficult for texts with mean lengths clearly below 5,000 words.The paper also reports that dates of writing are disputed and notes that a later proposed date for Acts makes common authorship more problematic.Source: §3 · pp. 4–5, Table 1 · p. 5
- Copyists may introduce variation by miscopying, omitting phrases, adding glosses, or inserting explanatory clauses.The paper notes that oral tradition may also stand behind some writings and gives Acts 8:37 as a verse some interpreters consider an addition.Source: §3 · p. 5
- Luke’s Gospel shares substantial wording with the other synoptic Gospels, so Matthew and Mark are not always used as comparison texts.The paper estimates that 41% of Luke’s Gospel is common with the other synoptic evangelists’ texts, 23% with Matthew, and 35% is unique.Source: §3 · p. 6, Table 2 · p. 6
- Greek has more complex functional-word morphology than English, including three genders, three numbers, and five grammatical cases.The paper notes seventeen possible forms for the definite determiner “the” and relatively free word order.Source: §3 · p. 6
4. Frequencies Considerations
Earlier arguments for common authorship relied mainly on qualitative similarities, while frequency differences have been cited as evidence against it. The section tests whether those differences are persuasive by comparing observed counts with expected frequency intervals and with differences between Dickens novels.
4. Frequencies Considerations
- Earlier studies offered mainly qualitative arguments for common authorship, while selected word-frequency differences challenged that hypothesis.Examples include ἐγένετο (69 occurrences in Luke, 54 in Acts), τε (9 in Luke, 151 in Acts), and καὶ (1,466 in Luke, 1,094 in Acts).Source: §4 · p. 7
- The test models the selected term’s count in Luke with a hypergeometric distribution, assuming its frequency is the same in Luke and the combined corpus.The corpus has 37,932 tokens, with 19,482 in Luke; for τε, the combined count is 160 and Luke’s count is 9.Source: §4 · pp. 7–8
- Because word use varies naturally, the analysis uses cumulative probabilities to set lower and upper count limits at selected confidence thresholds.The passage gives two-tailed thresholds such as 5%, 1%, or 0.5%.Source: §4 · p. 8
- In Luke and Acts, τε, οὖν, and εἰς occur too rarely in Luke under the equal-frequency hypothesis, while other selected terms also differ between the books.For τε, Luke has 9 occurrences against an expected interval of 70–95; for Acts’ lower-frequency examples, two are outside and two inside the confidence interval.Source: Table 3 · p. 7, §4 · p. 8
- Dickens’ two novels show that large differences in a few word frequencies do not by themselves establish different authorship.The paper cautions that a stylistic fingerprint cannot rest on a single or a few cases, and that word-frequency tests require care.Source: §4 · pp. 8–9, Table 4 · p. 9
5. Burrows’ Delta
The section applies Burrows’ Delta to Luke and Acts; nearest-text results favor a shared author, with higher probability estimates when Matthew and Mark are excluded.
5. Burrows’ Delta
- After earlier experiments did not provide an effective attribution method, the authors select Burrows’ Delta for a deeper quantitative test of common authorship.They frame the experiments as testing whether the shared-authorship hypothesis is confirmed or rejected.Source: §5 · p. 9
- Delta compares texts using frequent-word style markers, representing each term with a frequency-based Z score and averaging absolute score differences.The experiment varies the number of selected terms from 50 to 800.Source: §5 · pp. 9–10
- The authors propose applying softmin to Delta distances to estimate the probability that a query text was written by each candidate author.The constant c is fixed at 20 in the experiments to increase the dispersion of returned Delta values.Source: §5 · p. 10
- Matthew is always nearest to Luke’s Gospel, while Luke is the nearest text to Acts.For Acts, estimated probabilities increase through 300 terms and then plateau.Source: §5 · p. 11
- When Matthew and Mark are excluded, Acts becomes Luke’s closest text and Delta probabilities give higher confidence in Luke as author of both books.The authors caution that these probability estimates should be treated carefully because the study uses a limited number of texts.Source: Table 5 · p. 10, §5 · p. 11
6. Authorship Verification
The verification model compares 5,000-token chunks from Luke and Acts with impostor texts across feature sets and similarity or distance measures. It consistently supports common authorship, though results using fewer than 200 features are less reliable.
Authorship verification
- The verification procedure compares Luke-Acts chunks with impostor chunks, aggregates second-order similarity scores by their mean, and accepts authorship above the 80% threshold.The experiments use 1,000 iterations and randomly select 50% of the features for each comparison.Source: §6 · pp. 11–12
- Across the Table 7 experiments, the model confirms common authorship, but results with fewer than 200 features are less reliable.The table reports means over six chunk-level values; 1.0 indicates every chunk was assigned to Luke in all 1,000 iterations.Source: §6 · p. 13, Table 7 · p. 13
- When a chunk score falls below the 0.8 threshold, Table 7 marks the mean with an asterisk; in cited cases, only one of six values is below threshold.This occurs with cosine at 50, 75, 100, and 150 features, and also with MinMax, Dice, and Euclidean measures.Source: §6 · p. 13, Table 7 · p. 13
- With Manhattan distance and frequent n-grams, Table 8 always identifies Luke and Acts as written by the same person; MinMax experiments support the same conclusion.The n-gram representations use 3-, 4-, and 5-grams with feature sizes from 50 to 800.Source: §6 · p. 14, Table 8 · p. 15
- Truncation-word representations with 3–6 retained letters likewise detect common authorship across the reported feature sizes and in the MinMax appendix experiments.Each word is represented by its first n letters, with n set to 3, 4, 5, or 6.Source: §6 · p. 14
- The authors regard the conclusion as robust across parameter variations, while recommending at least 100 features for Greek because of its richer morphology and larger functional vocabulary.This recommendation qualifies the cautions about experiments with fewer than 200 features.Source: §6 · pp. 13–14
7. Stylistic Analysis
The stylistic analysis uses standardized term-frequency scores to identify words overused in Luke’s texts relative to other writers. It reports lexical and grammatical contrasts, while cautioning that isolated-word translations can be imprecise.
Stylistic analysis
- The Z score standardizes observed term frequency against its corpus-based expected frequency and binomial variance to assess whether a term is overused by an author.A positive score above the selected threshold indicates overuse; a score near zero indicates roughly proportional use.Source: §7 · p. 15
- Παῦλος appears 55 times in Luke against 9 occurrences in the other writers; its expected Luke count is 33.65, yielding a Z score of 7.28.With Lidstone smoothing at λ = 0.1, the example’s Z score becomes 7.37.Source: §7 · p. 16
- Luke overuses conjunctions, prepositions, third-person pronouns, εἶπεν, and names such as Paul, men, and Lord; other writers more often use Jesus, negations, and some prepositions.The comparison covers Luke against Matthew, Mark, John, Revelation’s ‘John,’ and Paul.Source: §7 · p. 16, Table 9 · p. 16
- The authors caution that translations of isolated words may be imprecise, and a two-word window may also be insufficient for precise bigram translation.The bigram analysis is reported as confirming the preceding table’s findings.Source: §7 · p. 17
- Luke more often uses past-tense εἶπεν, while the other writers prefer present-tense λέγει; Luke’s style also uses more verbs, whereas the others tend toward nouns.Luke’s genitive forms of the definite article also contrast with the more frequent nominative forms among the other authors.Source: §7 · p. 17
8. Conclusion
The study reports common authorship across both attribution strategies; it also cautions that stylistic inference is challenged by textual variation and limited comparable early-Christian corpora.
8. Conclusion
- Earlier studies used qualitative judgments or word-frequency differences, but a definitive conclusion had not been reached.Source: §8 · p. 18
- Burrows’ Delta indicated common authorship at all eleven feature-set sizes, and the verification model confirmed it across the tested representations, measures, and feature sizes.Source: §8 · p. 18
- The experiments suggest that reliable assessment requires 100 or more stylistic features and varying representation, feature-size, distance, and replication settings.Source: §8 · p. 18
- Literary dependence and copyist variation can interfere with stylistic signals, while limited early-Christian comparison texts constrain assessment.The authors note difficulty finding numerous securely attributed single-author texts longer than 10,000 words, on similar topics and from approximately the same period.Source: §8 · p. 18
Appendix
The appendix includes examples of synoptic-gospel similarity and a table of results using randomly selected features.
Appendix
- Table A.1 presents other examples of strong similarity between the synoptic gospels.Source: Table A.1 · p. 22
- Table A.2 gives an example of a clause added to Mark’s Gospel.Source: Table A.2 · p. 22
- Table A.3 reports MinMax similarity results for n-grams and trunc-n representations using a random selection of 50% of the features.Source: Table A.3 · p. 23