Who Wrote 1 Maccabees? A Computational Analysis

Computer analysis of the Greek text reveals that 1 Maccabees was rendered into Greek by a single translator — but composed from two distinct Hebrew source layers with strikingly different literary characters.

The Question

1 Maccabees tells the story of the Maccabean revolt against the Seleucid Empire (167–134 BCE), from the persecution under Antiochus IV Epiphanes through the restoration of the temple and the establishment of the Hasmonean dynasty. It is one of the most important historical documents from the Second Temple period.

But is it all by one person? The book contains strikingly different material: battlefield narratives, diplomatic letters to Rome and Sparta, a formal decree honoring Simon, and political correspondence with Antiochus VII. Some scholars have argued that these sections were added later by different hands. Others maintain that a single author compiled the whole book, weaving his own narrative together with documentary sources.

We can now test this computationally. Stylometry — the statistical analysis of writing style — measures patterns that authors produce unconsciously: how often they use common words like "and," "but," "the," and "in." These patterns are remarkably stable within an author's work and remarkably different between authors. By comparing these patterns across sections of 1 Maccabees and against other books from the same period, we can determine whether the text shows one stylistic signature or many.

What We Found

The analysis produced three major findings:

One translator rendered the entire book into Greek. The Greek text of 1 Maccabees maintains a unified stylistic signature throughout all sixteen chapters. A rolling analysis that slides a window across the entire book found no point where the macro-level style shifts toward a different hand. Even when tested against Judith — a book so stylistically similar that it is the hardest possible control — every window across all sixteen chapters stays closer to the 1 Maccabees core. This proves the Greek translation is the work of a single person.

But the Hebrew composition tells a different story. Function words measure the translator's habits — his unconscious patterns in Greek. To see the Hebrew author's habits, we need features that survive translation intact. One such feature is the Hebraistic "sons of X" construction (Hebrew בני + category noun), a biblical idiom that the translator rendered literally into Greek. This construction appears 19 times in chapters 1–9 and zero times in chapters 10–16, with no transition zone. At the rate observed in chapters 1–9, we would expect roughly 11 occurrences in the author's narrative prose in chapters 10–16. The probability of observing zero by chance is vanishingly small. This clean break — invisible to function-word analysis but unmistakable in the phrasal data — points to a compositional seam in the Hebrew original at the point where Judas Maccabeus dies (chapter 9).

The different sections use different registers, not different voices — at the Greek level. The diplomatic and decree sections (chapters 8, 12:1–23, 14–16) do differ from the narrative core (chapters 2–7, 9–11). A permutation test confirms this difference is statistically real, not random noise. But the words that distinguish these sections are topical — "Roman," "Spartan," "letter," "write," "high priest" — not the kind of function words (pronouns, prepositions, conjunctions) that change when a different translator picks up the pen.

The Greek reveals the author's source languages. This was the unexpected discovery. The ratio of δέ (de, "but/and") to καί (kai, "and") acts as a detector for whether a passage was translated from Hebrew or composed in Greek. Hebrew narrative uses paratactic chains — "and... and... and..." — which a translator renders as καί. Greek-composed text uses more δέ, a particle that signals contrast or continuation in a way Hebrew does not. By mapping this ratio across every section of 1 Maccabees, we can see the author's workshop: which sources he translated from Hebrew and which he copied from Greek originals.

The Source-Language Map

The δέ/καί ratio produces a gradient from Hebrew-translated to Greek-composed text. Displayed as a ratio (δέ occurrences per καί occurrence), the sections sort into three tiers:

Section	δέ/καί ratio	Source signal
Core A (ch 2–7)	0.028	Hebrew narrative
Core B (ch 9–11)	0.030	Hebrew narrative
Hasmonean decree (ch 14)	0.034	Hebrew (translated by author)
Roman excursus (ch 8)	0.041	Hebrew (translated by author)
Core C (12:24–13:42)	0.057	Hebrew with rising Greek influence
Spartan letters (12:1–23)	0.068	Greek correspondence
Assassination (ch 16)	0.084	Greek court records
Antiochus VII (ch 15)	0.140	Greek diplomatic archives

The core narrative chapters (2–7, 9–11) sit at the bottom of the scale: heavily paratactic — using short, chained "and... and... and..." sentences — with almost no δέ, consistent with close translation from a Hebrew Vorlage (source text). This confirms what scholars have long argued on other grounds — that 1 Maccabees was originally composed in Hebrew and survives only in Greek translation.¹

The surprise is what happens in the non-narrative sections. Two of the embedded documentary sections — chapter 8 (the Roman alliance) and chapter 14 (the Hasmonean decree) — have δέ/καί ratios indistinguishable from the Hebrew core. This means the original Hebrew compiler translated these foreign documents into Hebrew (or they were already circulating in Hebrew). Either way, they went through the same Hebrew-to-Greek translation pipeline as the rest of the book, producing a low δέ ratio.

At the other end, chapter 15 (Antiochus VII's correspondence) has a δέ/καί ratio five times higher than the core — a strong signal that the Greek translator preserved diplomatic language more or less as he found it. Chapter 16, the narrative of Simon's assassination, also shows elevated Greek influence, suggesting the compiler drew on Greek-language court records or official chronicles for events near his own time.

The Spartan letters (12:1–23) fall in between — moderate Greek elevation consistent with correspondence that passed through diplomatic channels in Greek.

This gradient is not a flaw in the book's composition. It reflects how the text passed through multiple hands: a Hebrew compiler who absorbed some foreign documents into his own idiom, and a Greek translator who rendered the whole into Septuagintal Greek — preserving the source-language fingerprints along the way.

Methodology

The analysis follows established computational stylometry methods: Savoy (2020) for distance metrics and vocabulary richness, Eder (2016) for rolling stylometry, and Bolt (2025) for feature adaptation to ancient texts.²³⁴ All data comes from the LXX-Rahlfs morphological database, which provides lemmatized word forms and grammatical tags for the entire Septuagint text.

Corpus

The test corpus consists of 1 Maccabees (18,292 words) divided into eleven internal segments based on content type:

Segment	Chapters	Words
Core A	2–7	7,116
Core B	9–11	4,792
Core C	12:24–13:42	1,407
Prologue	1:1–9	153
Persecution	1:10–64	909
Simon end	13:43–52	254
Roman excursus	ch 8	713
Spartan letters	12:1–23	406
Hasmonean decree	ch 14	1,104
Antiochus VII	ch 15	859
Assassination	ch 16	558

Four external control texts calibrate what "different author" looks like in this feature space:

Text	Words	Role
2 Maccabees	11,917	Different author, overlapping events
Judith	9,174	Different author, similar genre
Tobit (S)	7,233	Different author, different genre
3 Maccabees	5,110	Different author, different period

Feature Set

The analysis uses the top 100 most frequent lemmas (dictionary forms) across the combined corpus. Lemmas rather than surface forms are used because Greek inflection would otherwise split a single word into dozens of variants, diluting the signal.

Distance Metrics

Three complementary distance measures are computed for every pair of texts:

Burrows' Delta (Manhattan distance on z-scored frequency vectors). The standard metric in computational stylometry since Burrows (2002). Each lemma's frequency is converted to a z-score using the corpus-wide mean and standard deviation, then pairwise Manhattan distances are computed. This metric is robust to differences in text length and has been validated across dozens of languages and literary periods.⁵

Cosine Delta. Measures the angle between frequency vectors rather than their absolute difference. Less sensitive to outlier features; often preferred for shorter texts.

Kullback-Leibler Divergence (symmetric). An information-theoretic measure of how much one probability distribution differs from another. More sensitive to rare features than Delta.

Statistical Tests

Permutation test (1,000 iterations). The 1 Maccabees text is randomly split into two groups of the same size as the narrative core and the documentary sections, and the Delta between the random groups is computed. If the observed Core-vs-Documentary distance is larger than all random splits, the stylistic difference is statistically significant.

Bootstrap cluster stability (1,000 iterations). The feature matrix is resampled with replacement, and Ward's hierarchical clustering is performed on each resample. The percentage of resamples in which the Core segments cluster together measures how robust the Core's unity is.

Fisher's exact test with Bonferroni correction. Tests whether each individual function word differs significantly between the narrative core and the documentary sections, with correction for multiple testing.

Zeta analysis (Savoy 2020, Ch 7.1). Splits the narrative core and the documentary sections into ~500-word sub-documents and identifies words that consistently appear in one group but not the other. A scramble control (random split of Core into two halves) verifies that the method does not generate false positives.

Rolling stylometry (Eder 2016). A 2,000-word window slides across the entire concatenated text of 1 Maccabees in 500-word steps. Each window's frequency vector is compared to reference centroids (Core, 2 Maccabees, Judith), and the nearest centroid is recorded. This detects authorship transitions within a continuous text.

Results in Detail

Distance Matrix

The cosine delta matrix reveals the internal structure of the corpus:

Core cohesion. Core A and Core B have a cosine delta of 0.656 — the lowest pairwise distance in the entire matrix. Core C joins at 0.694 (to Core B) and 0.815 (to Core A). These distances are well below any cross-text comparison.

External separation. The four control texts form their own cluster: 2 Maccabees and 3 Maccabees pair at 0.251 (both are rhetorical Hellenistic Greek); Judith and Tobit pair at 0.543 (both are narrative historical fiction).

The overlap zone. Several internal segments — particularly the Spartan letters (1.159 from Core A), Antiochus VII (1.157), and the Prologue (1.095) — are more distant from Core A than Judith is (0.848). This overlap is expected: genre (battle narrative vs. diplomatic correspondence vs. opening summary) affects function-word distributions more than authorship does in translated texts. The critical test is rolling stylometry, which overcomes this limitation by measuring continuous stylistic proximity rather than static segment-to-segment distance.

Hierarchical Clustering

Ward's method clustering on cosine delta produces the following merge order:

1. 2 Maccabees + 3 Maccabees merge first (0.251) — the two rhetorical Greek texts find each other immediately.

2. Judith + Tobit merge next (0.543) — the two narrative historical fictions cluster together.

3. Core A + Core B merge (0.656) — the heart of 1 Maccabees.

4. Core C joins Core A+B (0.788) — the third Core segment is pulled in.

5. The remaining internal segments progressively join at higher distances, with the diplomatic sections (Spartan letters, Hasmonean decree, Antiochus VII) merging with each other before joining the Core cluster.

The dendrogram shows that 1 Maccabees segments do not scatter randomly among external texts. The Core forms a tight nucleus; the non-narrative sections form a secondary internal cluster; and only then do external texts merge in.

Rolling Stylometry

This is the decisive test. Three runs were performed:

Run A: Core vs. 2 Maccabees (2,000-word window, 500-word step). All 33 windows are closer to Core than to 2 Maccabees. No window in any chapter drifts toward the external text. This establishes the baseline: the entire book maintains one stylistic identity against a clearly different author.

Run B: Core vs. Judith (2,000-word window, 500-word step). This is the hard test. Judith is the closest external control — its static cosine delta to Core A (0.848) is lower than many internal segments. If any section of 1 Maccabees were genuinely by a different author, it might drift toward Judith in this test. Result: 33/33 windows closer to Core. Not a single window, in any chapter, is closer to Judith than to the 1 Maccabees Core centroid.

Run C: Three-way comparison (5,000-word window, 1,000-word step, Core + 2 Maccabees + Judith). All 14 windows identify Core as the nearest centroid.

The rolling stylometry detects no authorship transitions at the macro level. The stylistic signal is continuous from chapter 1 through chapter 16. An important caveat: the 2,000-word window is too large to isolate the smaller document insertions — the Spartan letters (406 words) and the Roman excursus (713 words) are each less than half a window, meaning they are smoothed over by the surrounding narrative. Rolling stylometry proves the unity of the macro-narrative; the δέ/καί ratio and Zeta analysis (above) are the appropriate tools for characterizing the micro-insertions.

Permutation Test

The observed Manhattan delta between the narrative core and the documentary segments is 0.482. In 1,000 random splits of the combined text into groups of the same size, the delta ranges from 0.248 to 0.390 (mean 0.318). The observed value exceeds all 1,000 random splits: p < 0.001.

This means the stylistic difference between the narrative core and the documentary sections is real — it is not an artifact of random variation. However, "different" does not mean "different author." The Zeta analysis (below) reveals what is driving the difference.

Zeta Contrastive Analysis

Zeta analysis identifies words that consistently appear in one group's sub-documents but not the other's.

Narrative core markers (words favoring Core over Documentary): Only 1 word exceeds the Z > 1.5 threshold — ἀπαίρω (apairō, "depart"). This is essentially no signal: the core has no distinctive vocabulary relative to the documentary sections.

Documentary markers (words favoring Documentary over Core): 20 words fall below Z < 0.5. Nearly all are topical: ἀντίγραφον (antigraphon, "copy of a letter"), γράφω (graphō, "write"), ῥωμαῖος (rhōmaios, "Roman"), ἀρχιερεύς (archiereus, "high priest"), ῥώμη (rhōmē, "Rome"), συμμαχία (symmachia, "alliance"), σπαρτιᾶται (spartiatai, "Spartans"), ἐπιστολή (epistolē, "letter"), δῆμος (dēmos, "people/assembly"), πρεσβευτής (presbeutēs, "ambassador").

These are the words of diplomacy and decree. They tell us the documentary sections are about different things, not by a different person.

Scramble control. When the Core text alone is randomly split in half and Zeta is rerun, the method finds 2 markers and 0 counter-markers — confirming it does not generate false positives from homogeneous text.

Fisher's Exact Test

Of 88 function words tested between the narrative core and the documentary sections, 17 show significance at p < 0.05 before correction. After Bonferroni correction for multiple testing (threshold p < 0.00057), only two survive:

• ὅπως (hopōs, "so that/in order that"): Core 0.8‰, Documentary 4.9‰
• δέ (de, "but/and"): Core 4.4‰, Documentary 9.6‰

Both are expected in Greek-composed (as opposed to Hebrew-translated) text. ὅπως introduces purpose clauses — complex, nested sentence structures (hypotaxis) — characteristic of Greek composition. δέ is the conjunction whose distribution maps source language, as discussed above.

Bootstrap Cluster Stability

Test	Segments	Stability
Core only (A + B + C)	3	94.9%
All internal segments	11	11.8%
Internal segments ≥ 500 words	8	34.8%

The Core segments cluster together in 94.9% of bootstrap resamples — exceptionally robust stability. The low all-internal stability (11.8%) reflects the instability of very small segments (the Prologue at 153 words, Simon's end at 254 words, the Spartan letters at 406 words) in clustering analysis, not genuine authorship diversity.

Vocabulary Richness

Vocabulary richness measures (TTR, Guiraud's R, Yule's K, Simpson's D) computed per segment show:

• Internal Yule's K range (segments ≥ 500 words): 412–552
• External Yule's K range: 302–408

The internal and external ranges overlap slightly, which is expected for texts in the same language, period, and broad genre. The key observation is that the Core segments (K = 429–487) cluster within a narrow band, while the external texts — particularly 3 Maccabees (K = 302) — are clearly distinct.

Two Substrates, One Translator

The combined evidence allows us to reconstruct the compositional history of 1 Maccabees with unusual specificity. The book as we have it is the work of a single Greek translator — but the Hebrew material he translated came from at least two distinct sources. This model has a precedent. Bickerman (1937) demonstrated that the author used two different chronological systems — the Macedonian Seleucid era for royal events, the Babylonian Seleucid era for Jewish religious events — inherited unconsciously from two types of sources.⁸ We find the same mechanism operating at the level of prose register.

The Heroic Substrate (Chapters 1–9)

The first half of 1 Maccabees reads like a biblical history. The author writes in "elegant biblical Hebrew" (Goldstein), modeling his narrative on Samuel and Chronicles. Mattathias is cast as a new Phineas (Numbers 25), Judas as a new David. The prose is saturated with Hebraistic constructions — "sons of Israel," "sons of foreigners," "sons of valor" — that would not occur in native Greek and reflect the idiom of the Hebrew original. Army sizes are drawn from Scripture: the 40,000 infantry and 7,000 cavalry at 3:39 come from 1 Chronicles 19:18 (David's Aramaean war), casting Judas's battle against Syria as a replay of David's.⁷

The author himself marks this material as a literary unit: "the rest of the acts of Judas, and of his wars, and of the noble deeds that he did" (9:22) — a heroic narrative composed in biblical register, probably drawing on oral and written Jewish traditions about the Maccabean revolt. The uniformly low δέ/καί ratio (0.028–0.030) across these chapters confirms that the Greek is close, literal translation from Hebrew.

The Documentary Substrate (Chapters 10–16)

At chapter 10, the character of the Hebrew original changes abruptly. The Hebraistic "sons of X" construction — 19 instances in chapters 1–9 — drops to zero. No transition zone, no occasional leakage. The typological army numbers disappear. Instead, the text becomes a documentary framework: over a quarter of the material (27.2%) consists of embedded diplomatic correspondence — royal letters, treaty texts, assembly decrees — with the author's narrative serving as connective tissue between them.

This material reads like an official administrative chronicle. (First Maccabees 16:24 explicitly confirms that the Hasmonean court kept exactly this kind of record, referring to the "chronicles of the high priesthood" of John Hyrcanus.) When Jonathan becomes the first Hasmonean High Priest (chapter 10), the narrative shifts from sacred history to political record-keeping. The δέ/καί ratio remains low through chapter 14 (indicating Hebrew composition), but rises sharply in chapters 15 (0.140) and 16 (0.084), where the translator preserved Greek diplomatic language more or less as he found it.

What Connects Them

A single Greek translator rendered both substrates into the same Septuagintal idiom. His function-word profile — the unconscious patterns of article usage, conjunction frequency, and preposition choice — is identical throughout. Rolling stylometry detects no transition at any point across all sixteen chapters: every 2,000-word window stays closer to the 1 Maccabees core than to any external control text.

The translator also handled embedded documents with editorial judgment. Some documents — the Roman treaty (chapter 8) and the Hasmonean decree (chapter 14) — were translated into Hebrew paratactic style, producing Greek whose δέ/καί ratio is indistinguishable from the core narrative. These may have already been circulating in Hebrew, or the translator (or a predecessor) may have rendered them himself. Other documents — the Spartan correspondence (12:1–23), the exchanges with Antiochus VII (chapter 15) — were preserved in their original Greek phrasing. The elevated δέ rates in these sections are the signature of text that never passed through Hebrew.

What We Cannot Determine

Whether the two Hebrew substrates were composed by the same person at different times, by an author and a continuator, or by two independent writers, the computational evidence cannot say. All three models produce the same data: unified Greek function words (same translator) with a sharp break in Hebrew phrasal patterns (different composition). The question of composite authorship has a long history: Ettelson (1925) argued for integrity; Schunck (1954) proposed source divisions based on content analysis; Neuhaus (1974) rebutted Schunck on literary grounds.⁹ Goldstein (1976) argued for single authorship on historical and literary grounds; Bar-Kochva (1989) refined the argument with detailed analysis of the military narratives.⁶⁷ None of these scholars had access to computational stylometric tools or the Hebraism distribution data presented here. The Hebraism evidence does not refute their conclusion of unity, but it introduces a compositional complexity — the sharp register break at chapter 10 — that traditional literary analysis did not detect.

Limitations

Stylometry has well-documented limitations that must be stated honestly.

Genre convergence. Function-word distributions are influenced by genre as well as authorship. Two different authors writing military siege narratives in translated Septuagintal Greek will produce similar profiles — this is why Judith (cosine delta 0.848 from Core A) is closer to the 1 Maccabees core than some internal segments are. Rolling stylometry mitigates this limitation by measuring continuous proximity rather than static distance, but it does not eliminate it entirely.

Small samples. Three segments — the Prologue (153 words), Simon's end (254 words), and the Spartan letters (406 words) — fall below the 500-word minimum recommended for reliable stylometric analysis. Results for these segments should be interpreted with caution.

Translator vs. author. Function-word stylometry (MFW, rolling analysis, Zeta) operates on the Greek text and therefore measures the translator's habits — not the Hebrew author's. Features that survive translation intact, such as the Hebraistic "sons of X" construction, are needed to probe the composition layer. This distinction is critical: the unified Greek profile proves one translator, but the 19-to-0 Hebraism break reveals a compositional seam in the Hebrew that function-word analysis cannot detect.

Indistinguishable substrate models. The computational evidence cannot distinguish between (a) a single Hebrew author who changed compositional mode at chapter 10, (b) an author and a continuator, or (c) a compiler-translator who wove together two pre-existing Hebrew documents. All three models produce the same data: unified Greek function words with a sharp break in Hebrew phrasal patterns. The historical and text-critical evidence must adjudicate between these models.

Confidence Assessment

Claim	Confidence	Key evidence
Single Greek translator for entire book	High	Rolling stylometry: 33/33 windows closer to Core; 94.9% bootstrap stability; no MFW transitions detected
Core narrative (ch 2–7, 9–11) is unified at Greek level	High	Core A–B cluster at 0.656 cosine delta; Zeta markers are topical only
Compositional seam in Hebrew at ch 10	High	19 Hebraistic constructions in ch 1–9, zero in ch 10–16; zero leakage in 5,772 words of narrative prose
δέ/καί ratio maps source language	High	Permutation p < 0.001; ch 8/14 Hebrew-like, ch 15/16 Greek-like; Fisher's exact confirms δέ as significant discriminator
Single Hebrew author for entire book	Uncertain	MFW evidence proves translator unity, not author unity; Hebraism break introduces doubt; cannot distinguish single author with mode shift from author + continuator

Conclusion

The Greek text of 1 Maccabees is the unified work of a single translator. The computational evidence — rolling stylometry, bootstrap cluster stability, and the absence of non-topical Zeta discriminators — all confirm that one person rendered the entire book into Greek.

But the Hebrew text he translated was not uniform. The Hebraistic "sons of X" construction — a biblical idiom that survives translation intact — reveals a sharp compositional seam at chapter 10. The first half of the book (chapters 1–9) is written in biblical register, modeling the Maccabean revolt on the wars of David and Joshua. The second half (chapters 10–16) drops this register entirely and reads as a documentary chronicle of Hasmonean political history. The break is absolute — 19 instances to zero, with no transition — and aligns with the disappearance of typological army numbers and the sudden rise of embedded diplomatic documents.

The δέ/καί source-language map adds a second dimension: a gradient from pure Hebrew translation (the core narrative) through semi-translated documents (the Roman treaty, the Hasmonean decree) to preserved Greek originals (the Spartan letters, the Antiochus VII correspondence, and the court records of chapters 15–16). Together, these two findings — the Hebraism seam and the source-language gradient — allow us to see the compositional layers of 1 Maccabees with a precision that traditional literary analysis alone could not achieve.

---

Bibliography

---